The Identification and Function of English Prosodic ... E. B.A. Liberal Arts Hampshire College, 2002

advertisement
The Identification and Function of English Prosodic Features
by
Mara E. Breen
B.A. Liberal Arts
Hampshire College, 2002
SUBMITTED TO THE DEPARTMENT OF BRAIN AND COGNITIVE SCIENCES IN
PARTIAL FULFILLMENT OF THE REQUIREMENTS FOR THE DEGREE OF
DOCTOR OF PHILOSOPHY IN COGNITIVE SCIENCE
AT THE
MASSACHUSETTS INSTITUTE OF TECHNOLOGY
SEPTEMBER 2007
02007 Massachusetts Institute of Technology.
All rights reserved
Signature of Author:
Department of Brain and Cognitive Sciences
August 20, 2007
Certified by:
Edward A. F. Gibson
Professor of Cognitive Sciences
Thesis Supervisor
Accepted by:
Matthew Wilson
Professor of Neurobiology
Chair, Department Graduate Committee
MASSACHUSETTS INSTITUTE.
OF TECHNOLOGY
AUG 2 8 2007
LIBRARIES
ARCHNEs
2
The Identification and Function of English Prosodic Features
by
Mara E. Breen
Submitted to the Department of Brain and Cognitive Sciences
on August 20, 2007 in Partial Fulfillment of the
Requirement for the Degree of Doctor of Philosophy in
Cognitve Science
ABSTRACT
This thesis contains three sets of studies designed to explore the identification and
function of prosodic features in English.
The first set of studies explores the identification of prosodic features using
prosodic annotation. We compared inter-rater agreement for two current prosodic
annotation schemes, ToBI (Silverman, et al., 1992) and RaP (Dilley & Brown, 2005)
which provide guidelines for the identification of English prosodic features. The studies
described here survey inter-rater agreement for both novice and expert raters in both
systems, and for both spontaneous and read speech. The results indicate high agreement
for both systems on binary classification, but only moderate agreement for categories
with more than two levels.
The second section explores an aspect of the function of prosody in determining
the propositional content of a sentence by investigating the relationship between syntactic
structure and intonational phrasing. The first study tests and refines a model designed to
predict the intonational phrasing of a sentence given the syntactic structure. In further
analysis, we demonstrate that-specific acoustic cues-word duration and the presence of
silence after a word, can give rise to the perception of intonational boundaries.
The final set of experiments explores the relationship between prosody and
information structure, and how this relationship is realized acoustically. In a series of
four experiments, we manipulated the information status of elements of declarative
sentences by varying the questions that preceded those sentences. We found that all of
the acoustic features we tested-duration, fO, and intensity-were utilized by speakers to
indicate the location of an accented element. However, speakers did not consistently
indicate differences in information status type (wide focus, new information, contrastive
information) with the acoustic features we investigated.
Thesis Supervisor: Edward A. F. Gibson
Title: Professor of Cognitive Sciences
Chapter 1
What is prosody?
Every native speaker of a language understands that a spoken message consists
not only of what is said, but also the way in which it is said. Prosody is the word used to
describe the characteristics of the acoustic signal which affect non-lexical meaning. It
describes the way in which words are grouped in speech, the relative prominence of
words in speech, and the overall tune of speech. It is comprised of psychological features
like pitch, quantity, and loudness, the combination of which give rise to the perception of
more complex prosodic features like stress (prominence), phrasing (grouping), and tonal
movement (intonation).
Focusing on non-lexical meaning, prosody is not concerned with describing
acoustic aspects of speech which determine a word's identity. For example, the
difference between the noun PERmit and the verb perMIT, is not considered a prosodic
one, even though it is the location of the prominence (stressed syllable, in this case)
which determines the difference. Prosody is concerned, rather, with acoustic features that
distinguish phrases and utterances from one another. For example, a sign occasionally
seen in men's restrooms, reads:
(1) We aim to please. You aim too, please.
The humor in the sign arises from contrast of the prominence and phrasing of the words
in the first sentence with the prominence and phrasing of the second. Therefore, the joke
arises from prosody.
To study prosody empirically, it is necessary to translate psychological features
into acoustic ones, which can be automatically extracted from speech and measured.
Therefore, prosodic investigations, including those contained in this thesis, use the
following acoustic measures, which have been shown to correspond to listeners'
perception. The acoustic correlate of pitch is fundamental frequency (FO), which is
measured in Hertz (Hz). The acoustic correlate of quantity is duration, which in the
current studies will be measured in milliseconds (ms). Finally, there are several acoustic
correlates of perceived loudness. Amplitude is an acoustic correlate of loudness, which is
a measure of sound pressure.It is measured in Newtons per square meter (N/m 2).
Intensity is another measure of sound pressure, which is computed as the root mean
square of the amplitude. Intensity is measured in decibels (dB). Energy is a measure of
loudness which accounts for the fact that longer sounds sound louder. As such, it is a
measure of the square of the amplitude times the duration of the sound.
The identification of prosodic features
We have already seen in (1) how prosody can determine meaning differences in
speech. To characterize the details of this relationship between prosody and meaning, we
need to decide what the critical features are, just as to study the relationship between
syntax and meaning we need to determine the important syntactic features. One approach
to studying this relationship is to impose categories on the acoustic spectrum, which is
what systems of prosodic annotation do.
4
Prosodic annotation
One way in which prosody researchers have addressed the problem of prosodic
feature identification is to find out what humans actually hear. This is achieved by
training human listeners to utilize coding schemes to tag speech with labels which
correspond to perceptual categories. In this way, coders can generate prosodicallyannotated corpora which can be used to ask questions about the function and meaning of
prosodic features.
The most widely-used system of prosodic annotation is the Tones and Break
Indices (ToBI) system of prosodic annotation (Silverman, et al., 1992), which was
developed by a group of speech researchers as the standard system of prosodic
annotation. However, in the time since the development of the ToBI system, several
limitations of ToBI have been suggested (Wightman, 2002). In response to these
limitations, prosody researchers have either changed the system to suit their own
purposes, or proposed alternative systems. The second section of this thesis will present
one such alternative proposal: the Rhythm and Pitch (RaP) system of annotation (Dilley
& Brown, 2005), and compare inter-rater reliability for both ToBI and RaP on a large
corpus of read and spontaneous speech.
The main challenge that prosodic annotation systems face is how to divide up the
continuous acoustic space into relevant categories. Should the categories be simply
perceptual in nature? Or should they, conversely, be determined by categories in
meaning? Studies of inter-rater agreement, such as the one presented in the second
chapter of this thesis, can attest to the effectiveness of a system in determining the
perceptual categories of prosody, but cannot speak to the effectiveness of a system's
determination of the meaning categories of intonation.
The function of prosody
As already indicated, prosody describes the relative prominence and grouping of
words in speech. The question this definition immediately raises, is: why? Why are
words grouped in particular ways? Why are some words more prominent than others? In
sum, what is the function of prosody for communication?
Multiple functions of prosody have been proposed and explored experimentally
which cover a wide spectrum of questions from many levels of language processing, from
word recognition, to sentence processing, to discourse processing, to questions about
emotion and affect in speech. The current discussion of the function of prosody will
focus on the levels of sentence and discourse processing, and will explore the way certain
prosodic features contribute to propositional meaning, and how other features contribute
to information structure.
The function ofphrasing
The first question we will address is: what is the function of prosodic phrasing?
Or, why are the words of utterances grouped into phrases? There are several functions
for phrasing, including production difficulty, pragmatic/semantic factors, and
performance factors. For example, (2) provides examples of how difficulty with lexical
access, in this case, of the low-frequency word "sextant," can results in particular word
groupings.
(2) I was looking for the / um // sextant.
5
Another reason for particular patterns of phrasing is exemplified in (3), which provides
an example of how pragmatic considerations can result in particular phrasings.
(3) The talk, as you know, starts at noon.
A final proposal about the function of phrasing is that it provides cues to the syntactic and
semantic structure of a sentence. This function is demonstrated in (4).
(4)
a.
The cop saw // the robber with the binoculars.
b.
The cop saw the robber // with the binoculars.
Upon hearing the rendition in (4a), a listener is most likely to interpret "with the
binoculars" as an adjunct of "the robber," indicating that the robber possesses the
binoculars. Conversely, in (4b), the listener is more likely to interpret "with the
binoculars" as an argument of "saw," indicating that the binoculars are the instrument
that the cop is using to see the robber.
This relationship between phrasing and sentence structure will be the focus of the
second section of this thesis. In addition, this section will explore which acoustic cues
give rise to the perception of phrase boundaries, by exploring whether acoustic cues to
boundaries correlate with results obtained from prosodic annotations of boundaries.
The function ofprominence
In addition to exploring the function of phrasing, this thesis will also explore the
function of prominence, and, specifically, the proposal patterns of prominence arise from
information structure. Information structure describes the role of utterances, or parts of
utterances, in relation to the wider discourse, and how these roles change over the course
of discourse.
Two wide categories have been proposed for information in discourse. They correspond
to information that is either new or old in the discourse. These two categories appear
under many different names, and with slightly different meanings, but old, or given
information, can be thought of that information which is under consideration by all of the
participants in the discourse. New, orfocused information, on the other hand, is,
according to Jackendoff, "the information in the sentence that is assumed by the speaker
not to be shared by him and the hearer" (1972, p.230).
In addition, there are two categories of focused information: new and contrastive. If an
entity is being introduced to the discourse for the first time, it is considered new. If an
entity is meant to contrast with or correct something already in the discourse, as is
Damon, in (5), then it is considered contrastive.
(5)
a. Did Bill fry an omelet?
(6)
b. No, Damon fried an omelet.
Several theorists have made explicit the relationship between certain types of
prominences and categories of information structure (Jackendoff, 1972; Steedman, 2000;
Pierrehumbert and Hirschberg,1990). For example, Pierrehumbert and Hirschberg argue
that a high pitch on a stressed syllable (a H* accent) indicates that the speaker is adding
the prominent entity to the discourse, and it is, therefore, new. Conversely, a steep rise to
a high pitch on a stressed syllable (a L+H* accent) indicates that the speaker means to
contrast the prominent entity with something already in the discourse.
6
The experiments in the third section of this thesis explore the relationship between
acoustics and meaning in terms of prominence relations. They will explore the acoustic
features that underlie the location of focus, and whether different acoustic features
underlie different types of focus, specifically, new and contrastive focus.
Chapter 2
The importance of prosodic factors in understanding and producing language is well
recognized by researchers studying many aspects of spoken language. However, the
relationship between the perception of prosodic events and underlying acoustic factors is
complex (e.g., Pierrehumbert, 1980; Choi, et al., 2005). Therefore, a useful and practical
means for investigating prosody has been through human annotation of prosodic
information. The current chapter will present the theoretical motivation and mechanics
of two such annotation systems, ToBI (Tones and Break Indices) (Silverman, K.,
Beckman, M., Pitrelli, J. Ostendorf, M. Wightman, C. Price, P., Pierrehumbert, J. &
Hirschberg, J., 1992)) and RaP (Rhythm and Pitch) (Dilley & Brown, 2005), which are
both based on phonological theory. In addition, we will present two large-scale intertranscriber agreement studies of both systems.
There are two approaches to prosodic annotation: phonetic and phonological.
Phonetic intonation systems are analogous to systems of phonetic transcription of
phonemes, like the International Phonetic Alphabet (IPA), in that they are concerned with
annotating the pitch events of speech without reference to categories or meaning, and
they intend the labels in the system to be applicable to any language. Phonological
systems, on the other hand, attempt to derive meaningful categories, which differ across
languages. The INTSINT (DeChristo & Hirst, 1987) and TILT systems (Taylor, 1992) are
phonetic; ToBI and RaP are phonological.
Several attempts at universally applicable phonological prosodic annotation
systems have been made (e.g., ToBI, and its related systems for other languages, and
RaP). The sections that follow will describe in detail the motivation and mechanics of
both the ToBI and RaP systems. In the sections on ToBI, we will first describe the
motivation and history of the ToBI system, and some of the theory on which the system
is based. Then, we will describe the components of a ToBI annotation and how a coder
applies the ToBI system to speech. Finally, we will describe recognized limitations of
the ToBI system, including a discussion of empirical research which fails to support the
categories embodied by the ToBI labels. Following discussion of the ToBI system, we
will present the motivation for the development of the RaP system, and the theory on
which it is based. Second, we willprovide details about the mechanics of RaP, and how
the RaP system is applied to speech. Finally, we will describe the ways in which
addresses the limitations of the ToBI system.
ToBI
The ToBI (Tones and Break Indices) system (Silverman, et al., 1992) was
developed in the early 1990s by a group of researchers from the fields of psychology,
phonology, and computer science. They intended ToBI to be a standard system which
could be used not only across different labs but also across disciplines, such that its use
would result in a large body of annotated speech. The following will describe the theory
on which the ToBI system is based, the inventory and mechanics of the annotation
system, and known limitations with the ToBI system.
ToBI History and Theory
A history of the development of the ToBI system for American English, and
subsequent variants for other languages can be found in Beckman, Hirschberg, and
Shattuck-Hufnagel (2005). The authors note that the ToBI conventions for pitch are
8
based on the theoretical work of Pierrehumbert and her colleagues (Pierrehumbert, 1980;
Beckman & Pierrehumbert, 1986, Pierrehumbert & Hirschberg, 1988). The ToBI
conventions for phrasing, or the prosodic grouping of words, are based on the work of
Price et al., (1991) and Wightman, Shattuck-Hufnagel, Ostendorf, & Price (1992). An
important tenet of the theory on which ToBI is based is the idea that tones in ToBI are
determined paradigmatically, meaning that high or low tones are indicated with reference
speaker pitch range, and not in relation to the local tones that immediately precede or
follow. This approach is hypothesized to result in high and low tones which are
comparable irrespective of the context in which they occur (Beckman, et al., 2005). In a
following section, this paradigmatic approach to tonal labeling will be contrasted with the
syntagmatic approach to tonal labeling embodied in the RaP system, in which tones are
labeled as high or low with specific reference to preceding tones (Dilley, 2005).
Label
Type
Metrical
RaP:
Intended to
capture
strong beat:
weak beat:
tier
no beat:
Tonal:
ToBITones
tier
RaP:
Pitch tier
Prominent sylls
H+!H*, !H*,
L+H*,
High: H*,
L+!H*
Low:
L*, L*+H,
L*+H
Non-prom sylls:
Major
boundary:
Minor
boundary:
Major
None
L-L%, H-H%, L-H%, H-L%,!H-L%
ToBI:
boundary:
Break
Index tier
RaP:
Rhythm
Minor
boundary:
3
)?
No boundary:
2, 1, 0
no label
Phrasal
ToBI
RaP
N/A
X, X?
x, x?
no label
!H-
L-,
4
tier
Table 1: Inventory of symbols and associated tiersfor ToBI andRaP
H*, L*, E*
H, L, E
H, L, E optionally used
to signal pitch change
)
9
ToBI Mechanics
A standard ToBI transcription consists of four tiers of symbolic labels which are
time-aligned with the speech signal: an orthographictier for labeling time-aligned text, a
tonal tier for labeling pitch events, a break index tier for labeling perceived disjuncture
between words and phrasing, and a miscellaneous tier for additional information. Recent
proposed modifications to the ToBI system include a fifth tier, termed an alternative (or
alt) tier (Brugos, Shattuck-Hufnagel, & Veilleux, 2006) where alternative choices for
tonal and break index labels may optionally be indicated. Determination of prosodic
labels is based both on a coder's perceptual impression of prosodic events, as well as on
the visual characteristics of the fundamental frequency (FO) contour. The inventory of
symbols used by the ToBI system is presented in Table 1. The tonal and break index tiers
form the core of a ToBI transcription, and will be described in detail in the following
sections.
Tonal Tier
An example of the ToBI tonal tier, and some of its associated labels, is provided
in Figure 11. The tonal tier enables the labeling of two kinds of information: pitch
accents and phrasal tones. Pitch accents in the ToBI system are indicated on syllables of
perceived prominence, i.e. syllables that are more perceptually salient than others in the
utterance. Pitch accents in ToBI are binary; a syllable is either accented or unaccented.
An important consequence of feature of the system is that the labeling of perceived
prominence in ToBI is also binary. This aspect of ToBI will be contrasted below with the
labeling or prominence in RaP, which allows for multiple levels of prominence labeling.
Pitch accented syllables are usually accompanied by a pitch excursion, which in
ToBI can be a movement to either a high or a low pitch in the speaker's range. Pitch is a
psychological concept, and is represented in prosodic annotation as its acoustic
counterpart, which is fundamental frequency (FO), measured in Hertz (Hz). There are a
total of eight pitch accent types, which are made up of high and low tones, and can be
simple, bitonal, or downstepped. The full inventory of ToBI pitch accent labels is
presented in Table 1. Simple pitch accents (H*, L*), are assigned to syllables where the
perceived prominence is associated with a single tone, which is either a local pitch
maximum (H*), or a local pitch minimum (L*). In addition, L* can indicate a
perceptually prominent syllable in a stretch of speech that is in the low part of the
speaker's range. Bitonal accents, (L+H*, L*+H, and H+!H*) are assigned to prominent
syllables where both a high and low tone combine to lend prominence to a syllable.
Specifically, the starred tone of a bitonal accent is associated with the stressed syllable,
while the unstarred tone leads or trails the starred tone on an unstressed syllable. Finally,
there are three "downstepped" variants of the simple and bitonal accents (!H*, L+!H* and
L*+!H) which are used when the pitch of a given high tone is lower than that of a
preceding high tone in the same phrase.
A total of eight tonal labels are also available for indicating hierarchical phrasal
information. Three phrase accents (H-, !H-, and L-), are used to indicate pitch movement
at a minor phrasal boundary, while two boundary tones (H% and L%) are used to
indicate pitch movement at a major intonational phrase boundary. Because the theory of
ToBI maintains that a major phrase always contains one or more minor phrases, a major
1 Soundfiles corresponding to examples in this paper can be found at:
http://tedlab.mit.edu/rap.html etc
phrase is always labeled with one of five phrase accent/boundary tone combinations (HH%, L-L%, H-L%, !H-L% and L-H%). Each of the preceding labels indicates
unidirectional rising or falling pitch movement, except L-H%, which generally indicates
bidirectional (falling-rising) movement.
Figure 1 provides an illustration of speech with the associated ToBI tonal labels,
and each label on the tonal tier will be described in turn. The H* on "Le-" indicates that
a perceptually prominent syllable produced with a high tone. The L- on "-gumes"
indicates a low tone associated with a minor (or intermediate) intonational phrase
boundary, which will be defined in the next section. The L* on "good" indicates a
perceptually prominent syllable in the low part of the speaker's range. The L* on "vi-"
also indicates a perceptually prominent syllable associated with a low tone, although here
it is associated with a pitch minimum, in that the pitch subsequently rises to the end of the
phrase, indicated by the "H-H%" on "vitamins."
CD00
200
150
Legumes are a good source of vitamins.
ToBI: tones
H
L* H-H%/o
L-
3 J1J1 J1
ToBI: breaks
RaP: rhythm
X
RaP: tones
H
1 1
4
xX
E
H
Figure 1. Example transcriptionof speech using the associatedToBI and RaP labels.
Break index Tier
A break index is a number from 0-4 which is assigned to the end of each word. In
general, this number indicates the perceived degree of disjuncture between words, with
the exception of the "2" label, which will be explained below. A "1"is used to indicate a
small degree of disjuncture, as found at phrase-medial word boundaries. A "0" indicates
a tight connection between words during fast speech or cliticization processes, e.g. didja
for didyou. A "3"indicates the perception of moderate disjuncture, which in the ToBI
system corresponds to a minor phrase boundary. A "4" indicates maximal disjuncture,
corresponding to a major phrase boundary.
There are two exceptions to the characterization of break indices as indicating
degree of perceived disjuncture. The first stems from the stipulation that particular break
indices must be used whenever a particular tonal label is indicated. In particular, a "3" or
"4" must be labeled whenever a phrase accent or boundary tone, respectively, is labeled
on the tonal tier, regardless of the perceived degree of disjuncture. Second, the break
index "2" is used to explicitly indicate a mismatch between tonal movement and
disjuncture information. Specifically, a "2" is reserved for word boundaries with "a
strong disjuncture marked by a pause or virtual pause, but with no tonal marks; i.e. a
well-formed tune continues across the juncture" or "a disjuncture that is weaker than
expected at what is tonally a clear intermediate or full intonation phrase boundary"
(Beckman & Hirschberg, 1994). Given this dual function, the "2" label can either
indicate (1) a large degree of disjuncture comparable to a 4, or (2) a small degree of
disjuncture comparable to a 1 (Pitrelli et al., 1994).
Once again, Figure 1 provides examples of ToBI break indices. The "3"
associated with "-gumes" indicates a minor boundary after "Legumes," and is
obligatorily accompanied by a phrase accent on the tone tier, "L-" in this case, indicating
a falling pitch at the boundary. The "4" associated with "vitamins" indicates a major
boundary. The label of "4" is obligatorily associated with a phrase accent/boundary tone
combination, "H-H%," indicating a tone that rises through the final syllable.
Limitations of ToBI
Although ToBI was designed to be the standard intonation annotation system for
the field, issues with the system have arisen throughout the years. The issues that will be
addressed in this section include (1) a discussion of experimental evidence that fails to
support the categories that are embodied in the ToBI system, (2) the way the ToBI system
has moved away from annotation approaches which instruct annotators to "label what
they hear" (Wightman, 2002), (3) low inter-transcriber reliability, and (4) the need for
labels indicating multiple levels of perceived prominence.
The most serious issue with the ToBI system is that it may not always reliably
capture the categories that speakers and listeners impose on intonation. Evidence for this
claim comes from a series of production and perception studies which demonstrate that
(a) speakers and listeners do not reliably produce or perceive the categories that ToBI
defines and (b) speakers and listeners do reliably perceive and produce more categories
than those defined by ToBI.
Evidence for the first claim-that speakers do not reliably perceive or produce the
categories embodied in the ToBI system-comes from experiments conducted by Bartels
& Kingston (1994), Ladd & Schepman (2003), and Watson, Tanenhaus, & Gunlogson
(2004), on the distinction that the ToBI system assumes between H* and L+H*. ToBI
assumes that these labels represent two categories of accents, both acoustically and
semantically. Acoustically, both L+H* and H* are accents aligned with peaks in the high
part of the speaker's range. Where they differ, however, is that H* is comprised of a
single high tone, whereas L+H* is comprised of two tones, a low target and a high target.
As a result, the rise to the peak of H* increases monotonically; conversely, the L+H* is
realized as "a high peak target on the accented syllable which is immediately preceded by
relatively sharp rise from a valley in the lowest part of the speaker's pitch range"
(Beckman & Hirschberg, 1994). Semantically, H* and L+H* are said to differ with
respect to the discourse status of the associated accented entity. Specifically,
Pierrehumbert and Hirschberg (1990) state that speakers use H* to mark new information
12
that should be added to the listener's discourse representation, while L+H* is used to
mark information that contrasts with something previously mentioned in the discourse.
To investigate the reality of an acoustic and semantic difference between H* and
L+H*, Bartels and Kingston (1994) synthesized a continuum of stimuli intended to vary
between H* and L+H* by independently manipulating four acoustic characteristics of the
target accent. They then presented sentences containing these synthesized accents to
naive listeners and asked them to make a meaning judgment with the assumption that
listeners would interpret entities accompanied by an H* as new to the discourse but
interpret entities accompanied by L+H* as contrastive with information already in the
discourse. The two main findings from their experiment were (a) that peak height, rather
than the shape of the rise from a low tone, distinguished H* from L+H* such that the
peak was higher for L+H* than for H*, and (b) that there was no clear evidence of a
categorical boundary between L+H* and H* in terms of meaning differences.
More evidence for the lack of a categorical acoustic distinction between H* and
from Ladd and Schepman (2003), who presented evidence which fails to
comes
L+H*
support the claim that L+H* is comprised of two tonal targets, compared to only one
target for H*. If H* is comprised of only one tonal target, then any dip in FO between
two H*s should not associate with a particular syllable. However, Ladd and Schepman
demonstrated that speakers will align a low pitch between two H* accents with the /n/ in
Jay Neeson or the /i/ in Jane Eason, depending on the intended production. In addition,
listeners were able to use speakers' alignment of the low pitch to disambiguate the
syllable membership of ambiguous consonants. These results suggest that the low pitch
between two H*s is in fact a tonal target, calling into question the categorical difference
between H* and L+H*, because ToBI assumes that it is only the latter accent in which
the L has phonological significance.
Finally, Watson, Tanenhaus, & Gunlogson (2004) explicitly tested the idea that
H* marks new material while L+H* marks contrastive material in an eye-tracking study.
Listeners heard the directions in (3), while interacting with the items in a computerized
display. The target item (e.g. camel/candle) in sentence 3c was crossed with the type of
accent aligned with it in a 2x2 design.
(3)
a. Click on the camel-and the dog.
b. Move the dog to the right of the square.
c. Now, move the camel/candle below the triangle.
L+H*/H*
Eye-tracking results demonstrated that listeners were more likely to look to the
contrastive referent (the camel) when they heard the L+H* accent than to the new
referent (the candle), suggesting that listeners quickly interpreted the L+H* accent as
marking contrastive information. In contrast, listeners did not look more quickly at the
new referent (the candle) when it was indicated with a H* accent. That is, they looked
with equal probability at the camel or the candle when it was produced with a H*,
indicating that H* is not preferentially treated as signaling new information. These
results suggest, therefore, that H* is not perceived as categorically different from L+H*.
In addition to empirical evidence that the ToBI system does not embody the
categories that speakers produce and listeners hear, it has also been criticized for the
13
grammatical constraints that it embodies, which are a result of the underlying theory
which ToBI assumes. Wightman (2002) notes that the development of the system has
led to a series of "linkage rules" whereby a label on one tier necessitates a particular label
on the other tier. For example, a label of '3' on the break index tier must be accompanied
by a phrase accent (L-, H-, or !H-) on the tonal tier. Similarly, a label of '4' on the break
index tier must be accompanied by a boundary tone (L-L%, H-H%, L-H%, H-L%, !HL%) on the tonal tier. Wightman argues that this interdependence between tiers leads
labelers away from labeling what they actually perceive.
The third noted limitation of the ToBI system is that, although previous studies
have noted high agreement for coarse binary comparisons like accent vs. non-accent,
labelers have demonstrated fairly low agreement for fine-grained distinctions such as
between accent types, and between phrase types. For example, although labelers in one
study exhibited 81% on the binary measure of the presence or absence of a pitch'accent,
their agreement on the type of pitch accent dropped to 64%, even when the least-agreedupon categories (i.e. H* vs. L+H*, !H* vs. L+!H*) were collapsed (Pitrelli, et al., 1994).
The final observed limitation of the ToBI system is the fact that it allows for only
a binary perceived prominence distinction. That is, syllables in ToBI annotations are
either pitch accented or unaccented. Conversely, throughout the history of intonational
theory, several researchers have proposed systems in which there are three or more levels
of prominence. For example, Halliday (1967) proposes a system in which there are
categories of stress and categories of accent. Accent is defined in terms of pitch, but
stress is defined with reference to rhythm. Empirically, there is evidence that three
categories of accent can be useful. Greenberg, Harvey, and Hitchcock (2002), for
example, found systematic differences between the pronunciation of words which were
produced with either a "heavy" accent, a "light" accent, or no accent. Finally, Beaver, et
al. (2007) have recently argued that second-occurrence focus, though not associated with
pitch changes, is still perceived as prominent, and therefore necessitates a way of being
indicated as prominent. As will be explained below, RaP instantiates a system with three
levels of prominence, in which metrical prominences are defined with reference to
rhythm, and pitch accented syllables are defined with reference to intonation.
One of the goals of the current chapter is to investigate whether a system without
the above limitations lends itself to higher inter-coder reliability. RaP addresses the
described limitations of ToBI in the following ways: First, tonal categories in RaP are
based on empirical investigations of the categories that speakers produce and listeners
perceive; Second, RaP allows for the independence of labels on different tiers; Finally,
RaP allows annotation of three levels of prominence, as opposed to only two in the ToBI
system.
14
RaP
The RaP (Rhythm and Pitch) system (Dilley & Brown, 2005) was developed to
meet the needs of the speech research community by building on experimental work and
theoretical advances that have taken place since the development of the ToBI system. The
following sections will describe the theory on which RaP is based, the mechanics of the
labeling system, and the way it addresses the limitations of ToBI.
RaP Theory
RaP is based on the work of Dilley (2005), which in turn draws heavily on the
work of Pierrehumbert and colleagues (Pierrehumbert, 1980; Beckman & Pierrehumbert,
1986; Goldsmith, 1976; and Liberman & Prince, 1977). Tonal labels in the RaP system
are based on the Tone Interval Theory of intonation and tone proposed by Dilley (2005).
Tone Interval Theory differs from the tonal theory upon which the ToBI system is based
in two important ways: First, whereas tones in ToBI are usually defined with reference to
the global pitch range of the speaker, the tones in RaP are defined with reference to the
preceding tone. Second, whereas ToBI postulates only high and low tones, RaP labels
reflect three types of tonal relations; a tone may be higher, lower, or at the same level as a
directly preceding tone.
Evidence that Tone Interval Theory provides a more accurate picture of the
categories underlying speaker's production comes from Dilley's thesis. In a series of
imitation experiments, Dilley demonstrated that the relative pitch level of one tone
(higher than, lower than, the same as a preceding tone) was perceived and produced
categorically. For example, where the ToBI system hypothesizes that H* and L*+H are
two categorically different accents, the RaP system predicts no categorical distinction
between these two accents.
Overview of the RaP Prosodic Transcription System
A RaP transcription is based on coders' auditory-perceptual impressions of
prosodic events. A visual display of the signal is considered an aid, not a requirement,
unlike ToBI. A transcription consists of four tiers of symbolic labels which are timealigned with the speech signal: a words tier for labeling time-aligned syllables, a rhythm
tier for labeling speech rhythm, a tonal index tier for labeling tonal information, and a
miscellaneoustier for additional information. In the following discussion we focus on
the rhythm and tonal tiers, which form the core of a RaP transcription.
Rhythm tier
The rhythm tier permits speech rhythm to be labeled by designating each syllable
as a metrical beat or not a beat. Several levels are distinguished. The label X is used to
indicate a syllable which is perceptually a strong metrical beat. The label x is used to
indicate a syllable which is perceptually a weak metrical beat. In addition, phrasal
boundaries are indicated at word boundaries with the following notation: ')' for a minor
phrase boundary; '))' for a full phrase boundary; no label for phrase-medial word
boundaries. RaP coders may indicate tonal labels in the tonal tier to account for
perceived pitch movement at phrasal boundaries, but they are not, as in the ToBI system,
obligated to do so.
Tonal tier
By separating tonal information from rhythmic information, the RaP system
makes it possible to distinguish syllables which are prominent due to the presence of a
15
pitch excursion (accented by means of pitch and thus pitch accents), versus syllables
which are prominent for rhythmic reasons. The tonal tier consists of labels indicating the
tonal targets that correspond to each syllable. All tonal events are indicated with a H
(high), L (low), or E (equal) label. In addition, RaP allows coders to indicate two sizes of
pitch change: small or large. Small pitch changes are indicated on syllables which incur a
pitch change of less than three semitones from the previous syllable, and are indicated
with the ! diacritic (e.g. !H, !L). Tonal targets which correspond to metrically prominent
(strong beat or weak beat) syllables (which are labeled with 'X' or 'x' in the metrical tier)
are called starred tones and are indicated with the * diacritic (e.g. H*). Furthermore,
tonal targets that occur on non-metrically prominent syllables (unstarred tones) that
precede or follow a starred tone by one syllable are labeled with a '+' (before or after the
tonal label depending on the location of the starred tone) indicating their association with
the adjacent starred tone (e.g. +H). Finally, as indicated above, tonal labels can be
indicated on phrase-final syllables which incur a pitch change.
In summary, there are two ways in which the RaP metrical and tonal tiers contrast
with those in ToBI. First, by allowing coders to indicate syllables which are metrically
prominent but do not incur a pitch change, RaP allows for three levels of prominence
labeling, as opposed to the two that ToBI allows (pitch accented vs. non-pitch accented).
Second, by assuming independence between phrasal and tonal labels, RaP allows labelers
to indicate perceived disjuncture without the perception of a pitch change.
Motivation for Study One
There are two independent, though equally important, motivations for Study One. One
is to conduct an inter-coder reliability study of the ToBI system, with an adequate
number of trained coders, a large corpus of speech, and the appropriate statistical
measures of agreement between coders. The second is to assess agreement of coders
using the RaP system.
Several studies of inter-coder reliability in the ToBI system have been published
since the development of the system; however, each has suffered from empirical
limitations. In the first study of ToBI inter-coder reliability (Pitrelli, et al., 1994), 26
coders applied the ToBI conventions to 489 words, taken from both read and spontaneous
speech corpora. Although this study employed many coders, the agreement results can
be questioned because of the-small amount of speech that each coder labeled, and
because the agreement metric used to compare coders did not take into account the
possibility of chance agreement. A more recent study by Syrdal & McGory (2000),
employed six coders to label 645 words. Although this study did take into account
chance agreement by employing the kappa statistic as described below, the agreement
numbers are also questionable because of the composition of the corpus of speech. First,
the corpus of speech is very small in comparison to the current study. Second, it is
comprised only of two speakers who read the same words, which suggest that the results
may not generalize to all speakers or to spontaneous speech. The final study of intercoder reliability in the ToBI system was conducted by Yoon, et al. (2005). Although this
study used appropriate statistical measures and a large corpus of spontaneous speech,
including 79 speakers and 1600 words, the speech was labeled by only two coders.
The studies presented in this paper were designed to address the limitations of
previous inter-coder reliability studies by employing (a) an adequate number of trained
coders, (b) a large corpus of varied speech, and (c) the appropriate statistical measures of
16
agreement between coders. In addition, they represent the first study of inter-coder
agreement for the RaP system.
Study One
The first study was designed to assess how transcribers with no previous prosodic
annotation experience would utilize both labeling systems. Undergraduates were trained
on both of the systems, and annotated fifty-five minutes of both read and spontaneous
speech in both systems. Their annotations were analyzed for agreement on multiple
prosodic characteristics, which will be described below.
Method
Participants
Five MIT undergraduates served as coders in the first study. Each coder committed to
spending a full summer on the project for which they received either course credit or
monetary compensation, at a rate of $8.75/hour for the duration of the project. Although
four of the coders had taken an introductory linguistics course, none had any knowledge
of prosody research, or any experience with prosodic annotation.
Corpus
There is some evidence that read speech is inherently different from speech which is
spontaneously produced. Hirschberg (1995), for example, notes that read and
spontaneous speech differ with respect to the intonational contours that speakers choose
for questions, and that read speech is faster than spontaneous speech. In order to ensure
that the results are applicable to diverse styles of speech, and to investigate differences in
coding agreement across different speech styles, materials for the present study were
drawn from two speech corpora: the Boston Radio News Corpus of read professional
news broadcast speech (Ostendorf, Price, & Shattuck-Hufnagel, 1995), and the
CallHome corpus of spontaneous nonprofessional speech from telephone conversations
(Linguistic Data Consortium, 1997). The amount of speech from each corpus which was
labeled in each system is shown in Table 1.
Table 1. Amount of speech (in minutes and syllables) from each corpus labeled in
each system, including number of coders per file. Speakers are the same for both ToBI
and RaP-annotated files.
System Corpus
ToBI
CallHome
BRNC
R
CallHome
BRNC
Total
Minutes
15.2
20.9
9.6
9.6
55.2
Sylles
3680
5939
2638
Coders/File
3.5
3.4
4.5
Unique Speakers
6
6
6
2889
15146
4.7
6
12
Coder training
Training and testing on prosodic systems occurred in three successive stages.
First, coders trained and were tested on the ToBI system, and then applied this system to
the speech corpus. Next, the coders trained and were tested on the RaP system, then
applied it to a subset of the corpus which had already been labeled with ToBI. More
details about training and labeling of the test materials are given below.
Trainingand testing of ToBI
Initial training on ToBI involved reading the associated manual and completing
the computerized exercises in Beckman and Ayers (1997), as well as receiving one-onone feedback from an expert coder (the author). In addition, all naYve coders participated
in weekly meetings with a group of four expert ToBI labelers throughout the course of
the project (the author, and three ToBI experts in the IT speech community). After two
weeks of initial training and feedback, the coders annotated one minute of read speech
from the BRNC corpus, which was not included in the ToBI test set. Feedback from two
expert coders (the author and Laura Dilley) was provided. Subsequently, the coders
annotated one-minute of spontaneous speech from the CallHome corpus, which was
again not included in the ToBI agreement analyses. Again, feedback from the two expert
coders was provided.
After these two feedback rounds, the coders labeled 90 seconds of speech
(approximately 60 seconds read speech, 30 seconds spontaneous), which were once again
separate from the ToBI test materials. The annotations were anonymously evaluated by
three expert coders (the author, Laura Dilley, and one of the IT experts) using the
following system: One or two points were deducted for each label with which the expert
mildly or moderately disagreed, respectively. Three points were deducted when a label
was strongly disagreed with and/or presented incorrect ToBI syntax. Experts also
employed a subjective grading system ranging from excellent (5) to poor (1), indicating
their overall impression of the labels.
Three coders received average grades of 4 or higher from all three expert
evaluators on both test files and began annotating the corpus. The other two coders
received average grades of 3 from the experts, and were instructed to go back through the
ToBI manual, paying attention to the labels they had misused in the test labels. After
another week of training, they too began corpus annotation.
Coders spent the next four weeks annotating 26.7 minutes of the corpus with the
ToBI system (11 spontaneous, 15.7 read). The order of files in the corpus was the same
for every coder, and pseudo-randomly determined so that coders would label
approximately equal amounts of radio and spontaneous speech, and not label speech from
the same speaker in adjacent files.
Following training and testing on the RaP system (as described below), coders
annotated the next 9.4 minutes of the corpus using ToBI (4.2 spontaneous, 5.2 read).
Inclusion of a second period of ToBI labeling permitted testing of the hypothesis that
higher agreement might result from more labeling experience in general, regardless of
the identity of the prosodic labeling system.
Trainingand testing of RaP
After the initial period of learning and applying ToBI, the coders spent two weeks
learning the RaP system. Coders were introduced to RaP using the guidelines laid out in
Dilley and Brown (2005)2. After an initial week of intensive group training with the
manual, coders annotated a one-minute passage of read speech, and received feedback on
their annotations from an expert RaP coder (Laura Dilley). Coders then labeled a oneminute passage of spontaneous speech and again received feedback from the expert
coder.
2Available
at http://tedlab.mit.edu/rap.html
After these two feedback rounds, the coders all labeled 60 seconds of speech
drawn from both the spontaneous and read corpora. The expert RaP coder gave the
novice coders quantitative and subjective scores for their annotations, as described above.
All coders received scores of "4" or above, and were cleared to begin annotating the
corpus according to the RaP conventions.
Coders spent the next four weeks annotating 19.2 minutes of the corpus using the
RaP system (9.6 spontaneous, 9.6 read). The files annotated with RaP were a subset of
the 26.7 minutes of the corpus labeled in the first four weeks of ToBI annotation.
Data analysis
Agreement metrics
All agreement analyses in the current study were designed to facilitate
comparison of current results to those of previous studies of ToBI inter-coder agreement.
Two measures of coder agreement were computed for the current study: one based on
raw agreement; the other correcting for chance agreement. Whereas in previous studies
of ToBI inter-annotator reliability accent labels are aligned with words (consistent with
the established Guidelines for ToBI Labeling (Beckman & Ayers, 1997)), accent
annotations in this study were aligned with syllables. This alignment scheme allowed us
to make direct comparisons between accent placements in both systems. However, it
should be noted that this scheme is different from previous studies of agreement, and
necessitates two types of raw agreement metrics, explained below.
Following the work of Pitrelli et al. (1994), our raw agreement measures are
based on the transcriber-pair-syllable(TPS) or transcriber-pair-word(TPW) depending
on the relevant comparison. For example, agreement on pitch accents was computed
syllable-by-syllable, while agreement for phrasing was computed word-by-word. We
would not want to compute agreement for phrasing on a syllable-by-syllable basis
because labelers would always trivially agree that there are not word-internal phrase
boundaries, thereby artificially inflating overall agreement. Agreement using the
transcriber-pair-syllable / word is computed as the total number of agreements between
pairs of coders on every syllable (or word) over the total number of possible agreements
between pairs of coders on every syllable (or word). To get raw agreement numbers for
each recording, each corresponding transcriber-syllable-pair or transcriber-word-pair is
checked. In this first step, all agreements are counted equally. Then a weighted mean of
the aggregate agreements for each recording is taken, weighted by the product of the
number of transcriber pairs and the number of units of comparison (i.e. words, syllables)
in the recording.
Because labels don't occur with equal frequency, guessing based on the frequency
of certain labels would allow labelers to produce relatively high agreement. For example,
Taylor & Sanders (1995) note that boundaries occur only 20% of the time. Therefore,
chance agreement on the annotation of boundaries is actually 80% (and not 50%), and
boundary agreement computed using the TPW method described above would be
artificially inflated. To correct for cases where chance agreement differs from 50%, the
current study employed the Kappa statistic. The standard Kappa statistic is given by the
following:
[1]
I = (Oa - Ea)/(l - Ea)
(1)
19
where Oa is the observed agreement and Ea is the expected agreement by chance, given
the statistical distribution of labels in the population. By convention, a kappa statistic of
.6 or higher indicates reliable agreement (Landis & Koch, 1977).
A kappa statistic was computed across for each comparison in the following way:
First, chance agreement was calculated for each transcriber pair over each recording,
based on the relative frequency of labels in the entire corpus. A weighted mean of
chance agreements was then computed by averaging chance agreement from each
recording, with each measure of chance agreement of each comparison type weighted by
the number of labels of that category in the recording. Finally, the kappa statistic for
each relation was derived from the weighted mean raw agreement and the weighted mean
chance agreement over the entire relation.
Agreement analyses - Metricalprominence
The first class of agreement concerned the presence and type of metrically
prominent syllables, and was computed only for RaP transcriptions, because it is only
applied to speech in the RaP system. Agreement on the presence of a metrically
prominent syllable consisted of a binary distinction. Coders agreed if they both labeled
(a) a strong or weak beat, or (b) no beat. Agreement on the type of metrical prominence
was based on a ternary distinction: Coders agreed if they both labeled (a) a strong beat,
(b) a weak beat, or (c) no beat.
Agreement analyses - Pitch accents
The second class of agreement concerned the presence and type of pitch accents.
Agreement on pitch accent presence was based on a binary distinction in both ToBI and
RaP. In ToBI, coders agreed if they both labeled (a) some variety of a pitch accent, or (b)
no accent; in RaP, coders agreed if they both labeled (a) some variety of a pitch accent, or
(b) no accent. We computed two quantifications of pitch accent type agreement for
ToBI, and one version for RaP. In ToBI, one quantification of pitch accent agreement
type was based on a ternary distinction: coders agreed if they both labeled (a) some
variety of a high accent, (b) some variety of low accent, or (c) no pitch accent. The
second quantification of ToBI pitch accent agreement involved a 6-way distinction where
all accents were treated as distinct, and compared to the lack of an accent. In this
analysis, all downstepped accents, except for H+!H*, were collapsed with their nondownstepped counterparts. -In RaP, coders agreed if they both labeled (a) some variety of
a high pitch accent (H*, !H*), (b) some variety of low pitch accent (L*, !L*), (c) an equal
pitch accent (E*), or (d) no pitch accent.
Agreement analyses - Phrasalboundaries
The third class of agreement concerned the presence and strength of phrasal
boundaries. Agreement on phrasal boundaries was only computed for word-final
syllables and was based on a binary distinction. In ToBI, coders agreed if they both
labeled (a) a minor or major phrase boundary, or (b) no boundary; in RaP, coders agreed
if they both labeled (a) a minor or major boundary or (b) no boundary. Agreement on the
type of phrasal boundary was limited to syllables on which one or more coders had
indicated a boundary and was based on a ternary distinction. In ToBI, coders agreed if
they both labeled (a) a minor boundary, (b) a major boundary, or (c) no boundary. In
RaP, coders agreed if they both labeled (a) a minor boundary, (b) a major boundary, or
(c) no boundary.
20
Results
TSP/TWP
ToBI RaP
kappa
ToBI RaP
Presence of beat (RaP only)
89%
0.78
Strength of beat (RaP only)
77%
10.61
Comparison
84% 0.69
78% 0.66
Presence of PA
Strength of PA
86%
84%
Strength of PA: All accents distinct (ToBI only)
75%
0.52
Presence of phrasal boundary
Strength of phrasal boundary
82%
79%
92% 0.50
86% 0.47
0.67
0.61
0.78
0.67
Table 2: Study One agreement results
The agreement results for Study One are presented in Table 2. Each agreement class
will be explained in turn.
Metrical prominence
The first class of agreement concerns the location and strength of metrical
prominences (beats), and applies only to the RaP labels. Agreement on the binary
distinction of beat presence was very high, as indicated by a TPS of 89%, and a kappa of
.78. Moreover, agreement on the ternary distinction of beat strength was high, indicated
by a TPS of 77% and a kappa of .61.
Pitch accents
The second class of agreement concerns the presence and type of pitch accent and
applies to labels in both RaP and ToBI. Agreement on the binary distinction of pitch
accent presence (present vs. absent) was equivalent for both labeling schemes, indicated
by a TPS of 86% and a kappa of .69 for ToBI, and a TPS of 84% and a kappa of .67 for
RaP. An examination of the three comparisons of pitch accent type revealed a correlation
between the number of pitch accent types being compared, and labeler agreement.
Specifically, the ternary accent distinction in ToBI (high, low, absent) resulted in a TPS
of 84%, and a kappa of .66. The four-way accent distinction in RaP (high, low, equal,
absent) resulted in a TPS of 78% and a kappa of .61. Finally, the six-way accent
distinction in ToBI (H*, L*, L+H*, L*+H, H+!H*, absent) resulted in a TPS of 75% and
a kappa of .52.
Phrase boundaries
The third class of agreement concerns the location and strength of boundaries in
both ToBI and RaP. Agreement on the presence of a phrasal boundary was a binary
distinction in both ToBI and RaP. This agreement was moderate for ToBI (TWP = 82%,
kappa= .50) and high for RaP (TWP = 92%, kappa = .78). Agreement on boundary
strength was, again, moderate for ToBI (TWP = 79%, kappa = .47) and high for RaP
(TWP = 86%, kappa =.67)
Discussion
In Study One, five students with no previous annotation experience labeled 36
minutes of speech with ToBI and 19 minutes of speech with RaP. A comparison of the
ToBI agreement numbers from the current study to those of previous examinations of
ToBI labeler agreement indicates that these students were proficient labelers. For
example, Yoon, et al. (2005) reported a kappa of .75 for a binary distinction of the
21
presence of a pitch accent, compared to the .70 reported here for both ToBI and RaP.
Second, they report a kappa of .51 for a ternary distinction of pitch accent type, compared
to the .67 kappa that we observed for the same comparison in ToBI.
The results also demonstrate that it is possible to achieve fairly high agreement
across many aspects of these two annotation systems. According to standard
interpretations of kappa, values over .60 indicate substantial agreement, and virtually all
of the agreement numbers obtained from the first study are above .60. In fact, there are
only two cases-the six-way accent distinction in ToBI and phrasal boundary
comparisons in ToBI-where agreement falls below .60. Even in these cases, agreement
is moderate. These data, therefore, indicate that annotators can achieve high agreement
in both annotation systems.
First, in the RaP system, labelers exhibited high agreement for the location and
strength of metrical prominences. This result indicates that the RaP system is a useful
alternative to ToBI in cases where information about speech rhythm is desired. In
addition, this result indicates that annotators can successfully identify multiple levels of
perceptual prominence using the RaP system.
Second, labelers exhibited high agreement on the presence and type of pitch
accent for both systems. Specifically, the kappa for pitch accent presence was identical
for both systems, suggesting that the labelers were using the same criteria across both
systems to decide on the location of pitch accents.
One of the stated goals of the current study was to ascertain whether coders
exhibited higher agreement for prominence in the RaP system, which allows the
differentiation of three levels of prominence, than in the ToBI system, which instantiates
only two levels of prominence. This claim cannot be answered by simply comparing
agreement on pitch accents across both systems, as it may be the case that the same
syllables were not labeled as pitch accents in both systems. A later analysis of the data
will explore the extent to which labelers agreed on labels of pitch accents across systems,
and what the pattern of disagreements looks like.
Finally, labelers were in higher agreement on the presence and type of phrasal
boundary in the RaP system than they were for the ToBI system. This finding may be a
result of the interference of ToBI's "linkage rules" with true perception. That is, because
ToBI demands that labelers annotate a tonal event wherever they annotate a phrasal
boundary, and vice versa, labelers may be more apt to label a phrasal boundary where
they don't hear one if they have labeled a tonal event, or more likely to not label a phrasal
boundary when they don't hear a tonal event. In the RaP system, conversely, there is no
interdependency between labels, so labelers can label using solely their perception of
both phrasal disjuncture and tonal events. Once again, an investigation of across-system
agreement can help to answer this question.
The finding of lower agreement for ToBI than for RaP on the presence and type
of phrasal boundary may indicate that RaP is a better system for phrasal labeling, perhaps
because it allows for the independence of labels across tiers. However, this result could
also be due to the order of labeling systems employed in the current study. That is,
because labelers did the majority of labeling with the ToBI system before they trained
and tested on the RaP system, their higher agreement in RaP may have simply been due
to an overall practice effect, rather than a difference between phrasal labeling
conventions of the two systems. We addressed this latter possibility by comparing
labeler agreement on the ToBI labels they completed before training and testing on the
22
RaP system with ToBI labels completed after training and testing on the RaP system.
The results of this analysis are presented in Table 3.
Comparison
Presence of PA
Strength of PA: High vs. Low
Strength of PA: All accents distinct
Presence of phrasal boundary
Strength of phrasal boundary
Before RaP
Kappa
TSP/TWP
0.69
86%
84%
0.66
75%
0.51
82%
0.48
80%
0.44
After RaP
Kappa
TSP/TWP
0.68
85%
83%
0.65
0.52
73%
86%
0.70
0.64
82%
Table 3: Agreement on ToBI corpus annotated before and after training on RaP
Comparing agreement on boundary labels in the ToBI system both before and
after training in the RaP system indicates that agreement on phrasing in the ToBI system
is indeed lower on labels completed before RaP training than after. However, it is not
simply the case that labeler were in higher agreement on their labels after they trained on
RaP. Rather, the difference in the kappa numbers in the two divisions of the ToBI corpus
reflects that fact that coders were using a wider variety of labels in the second division of
the corpus. Specifically, if the coders were using a larger distribution of labels, then
chance agreement decreased, and, therefore, the kappa scores increased.
Another way to ascertain whether or not RaP allows better agreement on phrasal
boundaries is to have coders who are experts in both systems label the same speech with
both systems, counterbalancing both the order of the presentation of the speech and the
order of the labeling systems. We took this approach in a second study of inter-annotator
reliability.
Study Two
Study Two was designed to address the limitations of the first study, which were
due to constraints on the availability of RaP training materials, and the availability of
expert labelers. For this study, we recruited four expert labelers to label a new corpus of
speech using both systems. Because all labelers were experts in both systems, we were
able to counterbalance the order of speech each coder labeled, as well as the order of
annotation. In this way, we could insure that any differences in agreement between
systems was not the result of the labelers being more proficient with either system, nor
the result of having labeled the same speech previously using another system.
Method
Participants
Four coders who were experts both in ToBI and RaP served as coders in the second
study. Two coders were undergraduates who had served as coders in the first study, and
continued to receive either course credit or monetary compensation, at a rate of
$8.75/hour, for the duration of the project. The other two coders were the author and
Laura Dilley.
Materials
The composition of the corpus for Study Two is presented in Table X. As indicated in
the table, speech was selected to ensure a balance between spontaneously produced
speech and read speech, and between male and female speakers. None of the material
used in the second study had been labeled as part of the first study.
FileName Duration (sec. Speaker
50
male
Spont 1
S ont2
42
male
46
female
Spont 3
Spont 4
44
female
Total
181
Radio 1
Radio 2
Radio 3
Radio 4
32
58
29
59
Total
178
male
female
female
male
1
Table X Duration(in seconds) and gender of speaker of speechfiles used in Study Two.
Procedure
Table X lists the order of speech files and coding systems used by each of the four
labelers in Study Two. Each labeler's order of speech files and systems was individually
determined to counterbalance the order in which the files were labeled in each system.
In order to ensure the highest possible agreement, coders labeled practice speech
files both before they began labeling the entire corpus, and before they switched labeling
systems. These practice files averaged 30 seconds. Each labeler then received feedback
on his/her labels from the Laura Dilley.
Labelers labeled individually, and never discussed their labels at any point during
the study.
Labeler 1
Labeler 2
Labeler 3
Labeler 4
FileName System FileName System FileName System FileName System
Practice
To I
pric
ToN30 Prt,~
S ont 1
Radio 1
Radio 2
Spont 2
ToBI
ToBI
ToBI
ToBI
S ont 1
Radio 1
Radio 2
ToBI
ToBI
ToBI
Ra
Practice
RaP
Piie -t-RaPU
Practice
ToBI
Practice,
ToBI
'-Practice
Spont 3
Radio 4
Radio 2
Spont 2
Radio 3
Spont 3
Radio 4
RaP
RaP
RaP
RaP
RaP
RaP
RaP
Sont 3
Radio 3
Radio
Spont 2
Radio 3
Spont 3
RaP
RaP
RaP
RaP
T0il
ToBI
ToBI
Spont 3
Radio 4
Radio 2
Sont 2
Radio 3
Spont 3
Radio 4
Radio 4
RaP
Spont 4
Practice; TOBL. Spont 4
Radio 3 ToBI t atit
Spont 3 ToBI Radio 3
Radio 4 ToBI Spont 3
ToBI
ToBI
RA
RaP
RaP
Spont 4
iitie
Spont
Radio
Radio
Spont
na
RAJce
RaP I Radio 3 1 RaP
1
1 RaP
RaP
2
2 TRaP
racti 3 T
ToBI
ToBI
ToBI
ToBI
ToBI
ToBI
ToBI
Spont 3
Radio 4
Spont 4
Practice
TOB
acl
a
Radio 3
Sont 3
Radio 4
Spont 4
Practice
Spont 1
Radio 1
ToBI
ToBI
ToBI
ToBI
RaP
RaP
RaP
Radio 2
anSpont 2
RaP y"fAwfee
Radio 3
Spont I
RaP
Spont 3
Radio 1
RaP
Radio 4
ToBI
RaP
RaP
RaP
RaP
RaP
",-TB1
ToBI
ToBI
Spont 4 IToBI
Radio 4
Spont 4
Spont 4 1RaP
RaP
RaPj
Radio 2
Spont 2
ToBI
ToBI
Table X File and system labeling orderfor allfour coders in Study Two.
Agreement Analyses in Study Two were calculated in the same way as they were
in Study One.
Results
The agreement results for Study Two are presented in Table X. Each agreement class
will be explained in turn.
The first class of agreement concerns the location and strength of metrical
prominences (beats), and applies only to the RaP labels. As in the first study, agreement
on the binary distinction of beat presence was very high, as indicated by a TPS of 89%,
and a kappa of .78. Moreover, agreement on the ternary distinction of beat strength was
high, indicated by a TPS of 79% and a kappa of .63.
The second class of agreement concerns the presence and type of pitch accent and
applies to labels in both RaP and ToBI. Agreement on the binary distinction of pitch
accent presence (present vs. absent) was slightly higher for the ToBI system, indicated by
a TPS of 89% and a kappa of .76 for ToBI, and a TPS of 85% and a kappa of .66 for RaP.
In contrast with the results of the first study, there was no correlation between the number
of pitch accent types being compared, and labeler agreement. Specifically, the ternary
accent distinction in ToBI (high, low, absent) resulted in a TPS of 86%, and a kappa of
.71. The four-way accent distinction in RaP (high, low, equal, absent) resulted in a TPS
of 80% and a kappa of .60. Finally, the six-way accent distinction in ToBI (H*, L*,
L+H*, L*+H, H+!H*, absent) resulted in a TPS of 78% and a kappa of .58.
Comparison
TSP/TWP
kappa
ToBI RaP
ToBI
0.78
0.63
89%
79%
Presence of beat (RaP only)
Strength of beat (RaP only)
Presence of PA
Strength of PA
89%
86%
Strength of PA: All accents distinct (ToBI only)
78%
Presence of phrasal boundary
Strength of phrasal boundary
91%
87%
85%
80%
RaP
0.76
0.71
0.66
0.60
0.58
90%
85%
0.76
0.68
0.76
0.67
Table X Agreement resultsfrom Study Two.
The third class of agreement concerns the location and strength of boundaries in
both ToBI and RaP. Agreement on the presence of a phrasal boundary was a binary
distinction in both ToBI and RaP. This agreement was similar for ToBI (TPW = 91%,
kappa = .76) and RaP (TPW = 90%, kappa = .76).
Agreement on the ternary distinction
of boundary strength (high, low, none) was also comparable for both systems, as
indicated by a TPW of 87% and a kappa of .68 for ToBI and a TPW of 85% and a kappa
of .67 for RaP.
25
Discussion
In a second study of inter-annotator reliability of both the ToBI and RaP
annotation systems, we employed four expert labelers to apply each system to the same
six minutes of speech in each system. In contrast to the first study, the order of both
speech files and annotation systems was counter-balanced across labelers to control for
differences in agreement due to better proficiency in one system over the other, or any
possible advantage obtained from labeling the same speech twice. Broadly speaking,
agreement numbers for the second study were similar to those from the first study. Once
again, virtually all kappa values indicated substantial agreement in that they exceeded .60
As in Study One, agreement was very high for the location and strength of
metrical prominences in the RaP system. As is Study One, agreement on the presence and
tone of pitch accent was high for ToBI and RaP. The advantage we observed in the first
study for boundary agreement and strength in the RaP system was not apparent in the
second study.
General Discussion
The current study was conducted both to provide a large-scale investigation of the
inter-coder reliability of the ToBI system of prosodic annotation, and to ascertain the
inter-coder reliability of a new system of prosodic annotation, RaP, which was designed
to address the known limitations of the ToBI system. The two annotation systems are
intended to capture several different prosodic categories. Specifically, ToBI endeavors to
capture the presence and strength of intonational boundaries (i.e. perceptual breaks) and
the presence and tonal characteristics of pitch accents, which serve to give certain
syllables perceived prominence. RaP aims to capture the same information about
boundaries and accents and, in addition, allows for the coding of speech rhythm.
The results of the first study of rater reliability showed high agreement on every
tested comparison for both systems among naYve coders for both ToBI and RaP. First,
they demonstrated moderate-to-high agreement on presence and tone of phrasal
boundary, though agreement was higher on both of these characteristics for the RaP
system. As for pitch accent, agreement correlated with number of accent categories; that
is, agreement dropped steadily as agreement went from a 2-way distinction to a five-way
distinction. Finally, the results demonstrated high agreement in RaP on the coding of
metrical prominence.
A further result from the first study was the observation that coders demonstrated
higher agreement on the presence and strength of phrasal boundaries in ToBI in the
portion of the corpus that they labeled with ToBI after they had learned the RaP system.
The kappa results suggest that the labelers were using a wider distribution of boundary
labels in the annotations they completed after learning RaP.
The results of the second study replicated the pattern of results of the first study;
namely, high agreement for both systems for the categories of boundary presence and
strength, high agreement for both systems for pitch accent presence and type, which
decreased with the number of pitch accent categories, and high agreement for RaP on
pitch rhythm. Moreover, the results of the second study, which was done using the
annotations of expert coders, suggest that the coders in the first study may have had
higher agreement on RaP categories as a result of leaming the RaP system after the ToBI
system.
26
Overall, the results of the two studies indicate that coders can achieve high
agreement on both systems, and that prosodic annotation can be considered a valuable
tool for use in the study of prosody. Furthermore, they demonstrate that RaP is a viable
alternative to ToBI for the annotation of large speech corpora, especially when speech
rhythm information, or information about multiple levels of prominence, is desired.
Future analyses of the corpus annotated for the current study will be used to
explore questions about whether the differences in the labeling inventories of the ToBI
and RaP systems impact agreement across the two systems.
Chapter 3
The focus of the current chapter is on understanding how the syntax of a sentence can be
used to predict how a speaker will divide that sentence into intonational phrases. An old
joke, illustrated in (1), serves as an example of how different intended syntactic structures
can be realized with different intonational phrasing:
(1)
Woman without her man is nothing.
Woman // without her man // is nothing.
Woman! Without her // man is nothing.
In (la), the sentence is disambiguated by the insertion of a break after 'man,' such that 'is
nothing' must refer to 'Woman.' Conversely, in (Ib), the break is placed before 'man,'
thereby forcing 'is nothing' to modify 'man.' The written breaks correspond to points of
disjuncture in the sentence-places where one intonational phrase ends, and another
begins-and so are referred to as intonational boundaries. These two prosodic parses are
the direct result of two different syntactic structures, corresponding to two different
sentence meanings. The goal of the current research is to determine the nature of the
relationship between syntactic structure and intonational boundary placement.
The location of an intonational boundary, such as that which would be produced
after "man" in (1a), is indicated by the presence of several acoustic characteristics. A
highly salient cue to a boundary location is the lengthening of the final syllable preceding
a possible boundary location (Wightman, et al., 1992; Selkirk, 1984; Schafer, et al., 2000;
Kraljic & Brennan, 2005; Snedeker & Trueswell, 2003, Choi, et al., 2005). Characteristic
pitch changes accompany this lengthening. The most common of these changes are a
falling tone (as in that which ends a declarative sentence) or a rising tone (as in the
default yes-no question contour) (Pierrehumbert, 1980). Other cues, such as silence
between words (Wightman, et al., 1992), and the speaker's voice quality (Choi, et al.,
2005), can also cue the presence of a boundary, though the systematic contribution of
each cue has not been explored empirically.
There is some debate in the literature about what function the production of
intonational phrasing serves. Previous studies have shown that speakers often
disambiguate attachment ambiguities with prosody (Snedeker & Trueswell, 2003;
Schafer, et al., 2000; Kralji'c & Brennan, 2005), and that listeners can use such
information to correctly interpret syntactically ambiguous structures (Snedeker &
Trueswell, 2003; Kjelgaard & Speer, 1999; Carlson, et al., 2001; Speer, et al., 1996;
Price, et al., 1991). However, a question remains as to whether speakers produce
boundaries as a communicative cue for the benefit of the listener, or, alternatively, as a
by-product of speech production constraints. Data supporting the first possibility comes
from Snedeker and Trueswell (2003), who found that speakers only prosodically
disambiguated prepositional phrase attachments in globally ambiguous sentences like
Tap thefrog with thefeather if they (the speakers) were aware of the ambiguity. This
result suggests that the mere act of planning the syntactic structure of the sentence for the
purposes of producing it did not induce speakers to reflect the syntactic structure in their
intonational phrasing, indicating that some boundaries are not produced as a normal byproduct of speech planning. In contrast to Snedeker and Trueswell's results, Schafer et
al. (2000) showed that speakers produced disambiguating prosody early in a sentence
even when the intended meaning was lexically disambiguated later in the sentence (When
28
the triangle moves the square / it... vs. When the triangle moves // the square will...).
This result suggests that speakers produce some boundary cues even when such cues are
not required for the listener to arrive at the correct syntactic parse of the sentence, and
therefore, the production of intonational boundaries is a normal part of speech planning.
Additional support for the position that boundaries are a result of production constraints
comes from recent work by Kraljic and Brennan (2005), who showed that speakers
consistently produced boundary information whether or not such information would
enable more effective communication with a listener. In their first experiment, Kraljic
and Brennan employed a referential communication task in which speakers produced
syntactically globally ambiguous instructions like (2) for listeners, who had to move
objects around a display.
(2) Put the dog in the basket on the star.
The referential situation either supported both meanings of the sentence, or only one.
The placement of intonational boundaries can disambiguate the appropriate attachment as
follows: If "in the basket" is meant to indicate the location of the dog (the modifier
attachment), then speakers can do so with the inclusion of a boundary after "basket," as
indicated in (2a). If, conversely, "in the basket" is meant to be the location for the
movement (the goal attachment), speakers can indicate this meaning with a boundary
after "dog" as in (2b).
(2a) Put the dog in the basket // on the star.
(2b) Put the dog // in the basket on the star.
If prosody is produced as a cue for listeners, then speakers should be more likely
to produce disambiguating prosody when the situation has not been disambiguated for the
listener. Contrary to this hypothesis, Kraljic and Brennan observed that speakers
disambiguated their instructions with prosody even when the referential situation had
been disambiguated for them, in cases where the listener did not need an additional
prosodic cue to correctly interpret the instruction. Kraljic and Brennan conclude that
boundary production is a result of the speaker's processing of the specific syntactic
structure s/he is producing. This statement leads directly to the question we will be
addressing in this paper: What is the relationship between syntactic structure and
speakers' production of intonational boundaries?
The Left-Right Boundary Hypothesis.
Over the past several decades, researchers have attempted to use syntactic structure to
predict the placement of intonational boundaries. Gee and Grosjean (1983), Cooper and
Paccia-Cooper, (1980) and Ferreira (1988) have all presented models that attempt to
account for boundary placement in terms of syntactic structure. Roughly speaking, in
each of these models, longer constituents, syntactic boundaries, and major syntactic
categories like matrix subjects and verbs correspond with a higher probability of an
intonational boundary, although the weighting of each of these factors differ across
models.
Watson and Gibson (2004) (W&G) tested the predictions of Cooper and PacciaCooper's, Gee and Grosjean's and Ferreira's models by having naYve speakers produce a
series of sentence structures varying in length and syntactic complexity.
In order to encourage natural, fluent productions, W&G employed an interactive
communication task, where speakers were paired with listeners and encouraged to
produce normal fluent speech. First, speakers read a target sentence silently. The speakers
then produced the target sentences aloud for listeners, knowing that the listeners would
have to answer comprehension questions about what the speakers produced. In this way,
speakers were aware that they were conveying information to listeners, and so would be
more likely to produce the sentences with the cues they would use in normal, interactive,
language. In addition, speakers were familiar with the material they were producing, and
would be less likely to produce disfluencies, an outcome that W&G intended in their
choice of task.
W&G identified several aspects of Gee and Grosjean's, Cooper and Paccia-Cooper's and
Ferreira's models which enabled successful boundary placement predictions with respect
to their data set. First, the size of the most recently completed constituent was a good
predictor of boundary occurrence, regardless of the constituent's position in a syntactic
tree of the sentence. Assuming a model whereby boundary production behavior is driven
by the needs of a speaker, as supported by Kraljic and Brennan (2005), W&G suggested
that this effect is due to a 'refractory period' in which the speaker must recover from the
expenditure of resources involved in producing a sentence. A second generalization that
W&G observed is that the three models all predict more boundaries to occur before the
production of longer constituents. Again, W&G interpreted this effect as a result of
speaker needs; specifically, that speakers take more time to plan longer upcoming
constituents. This possibility is supported by evidence from Ferreira (1991) that speakers
take longer to initiate speaking when a sentence has a longer utterance-initial NP, and
from studies by Wheeldon and Lahiri (1997) showing that speaker's latency to begin
producing a sentence increases as a function of the number of phonological words in the
sentence.
Based on their observations of the pervasiveness of the two preceding phenomena, W&G
proposed a model of boundary production based on speaker resources. They
hypothesized that speakers place boundaries where they do to facilitate recovery and
planning. Specifically, W&G suggested that speakers use boundaries to recover from the
expenditure of resources used in producing previous parts of an utterance, and to plan
upcoming parts of an utterance. As such, they calculated the probability of a boundary at
a certain point (which they subsequently call the "boundary weight") as a function of the
size of the preceding material (Left-hand Size-LHS) and the size of the upcoming
material (Right-hand Size-RHS). It should be noted that, although in their 2004 paper
W&G refer to the material that precedes a candidate boundary location and the upcoming
sentence material as the LHS and RHS, respectively, we will, in this paper, emphasize the
relationship of these two components to the speaker's production process, and so will
refer to them throughout as Recovery and Planning, respectively.
Size in this model is quantified as the number of phonological phrases in a region. A
phonological phrase is defined as a lexical head and all the maximal projections on the
head's left-hand side (Nespor and Vogel, 1986). Specifically, a phonological phrase
contains a head (noun or verb) and all the material that comes between that head and the
preceding one, including function words, pre-nominal adjectives, and pre-verbal adverbs.
The following examples, from Bachenko and Fitzpatrick (1990) demonstrates how a
sentence is divided into phonological phrases:
a. A British expedition I launched Ithe first serious attempt.
30
b. We saw Ia sudden light I spring up Iamong the trees.
W&G chose phonological phrase boundaries as the most likely candidates for
intonational boundaries as several researchers had noted that, in fluent utterances,
speakers rarely place boundaries within phonological phrases (Nespor & Vogel, 1986;
Gee and Grosjean, 1983). This phonological phrase constraint greatly limits possible
boundary points within a sentence.
In addition to the size constraint on boundary production defined in terms of phonolgical
phrases, W&G hypothesized that boundaries would not occur between heads and their
arguments. This constraint was motivated by the work of Selkirk (1984), who proposed
the Sense-Unit Condition constraint on boundary production. The Sense-Unit Condition
stipulates that constituents within an intonational boundary must together form a sense
unit, meaning that each constituent within the intonational phrase must participate in
either a modifier-head or argument-head relationship with another constituent in the
phrase. An example is given in (4).
(4)
a. The mayor of Boston was drunk again.
b. The mayor I of Boston was drunk again.
(4a) provides an allowable phrasing of the sentence "The mayor of Boston was drunk
again," but (4b) is disallowed by the Sense-Unit Condition because the two phonological
phrases "of Boston" and "was drunk again" do not form a sense unit. One does not
depend on or modify the other; rather, they both depend on "The mayor."
As a first attempt at quantifying semantic-relatedness according to the Sense-Unit
Condition, W& G constrained the Recovery and Planning weights as follows: First, the
Recovery weight of a constituent immediately preceding a possible boundary location is
zero if the following constituent is a dependent of the preceding constituent; Second, the
Planning weight of a constituent following a possible boundary location is zero if that
constituent is an argument of the immediately preceding constituent.
W&G operationalized their predictions with the Left/Right Constituent Boundary
Hypothesis (LRB), which is stated in (3):
Left Constituent / Right Constituent Boundary (LRB) Hypothesis: Two
(3)
independent factors are summed to predict the likelihood of an intonational boundary at a
phonological phrase boundary in production.
Recovery (LHS): The number of phonological phrases in the largest syntactic constituent
that has just been completed. A constituent is completed if it has no rightward
dependents.
Planning (RHS): The number of phonological phrases in the largest upcoming syntactic
constituent if it is not an argument of the most recently processed constituent.
In more recent work, Watson, Breen, and Gibson (2006) investigated the nature of
the semantic-relatedness constraint on boundary production. They found that rather than
there being a restriction on the placement of boundaries between heads and their
immediately adjacent arguments, a better model fit is obtained if boundaries are
disallowed between heads and their immediately adjacent obligatory arguments.
Specifically, they found that although speakers can be induced to place boundaries
between nouns and their non-obligatory arguments if the argument is long (e.g. The
31
reporter's arrival // at the crash of the subway vs. The reporter's arrival at the crash),
speakers virtually never place boundaries between verbs and their obligatory arguments,
even when the argument is long (e.g. The reporter arrived at the crash of the subway).
The current study does not provide evidence in support of a particular formulation of a
semantic-relatedness constraint, as the obligatoriness of arguments is not systematically
manipulated in the current study. However, we will assume the most recent published
formulation of the LRB (indicated in 4), in which boundaries are disallowed between
heads and their adjacent obligatory arguments.
Left Constituent / Right Constituent Boundary (LRB) Hypothesis: Two independent
factors are summed to predict the likelihood of an intonational boundary at a
phonological phrase boundary in production.
1) Recovery (LHS): The number of phonological phrases in the largest syntactic
constituent that has just been completed. A constituent is completed if it has no
obligatory rightward dependents.
2) Planning (RHS): The number of phonological phrases in the largest upcoming
syntactic constituent if it is not an obligatory argument of the most recently processed
constituent.
According to the original definition in (3), the LRB accounted for a significant amount of
variance in boundary placement (r2 =. 74, N = 85, p<. 001). When compared to the
predictions made by the three other algorithms, the LRB was able to make similarly
accurate predictions, using fewer parameters. It should be noted that W&G's regression
analysis is somewhat inflated in two ways: First, it was computed on all word boundaries,
including places where boundaries virtually never occur in fluent speech (e.g. between
articles and nouns) (Watson & Gibson, 2004). Indeed, when the same regression was
performed only on word boundaries that coincided with phonological phrase boundaries
(the locations where the LRB predicts boundaries will occur), the LRB accounted for a
diminished, though still significant, amount of the variance in boundary production (r2
.55, N = 41, p<. 001). Second, W&G computed their regression analyses using the
average boundary proportion of each of the eight constructions. Here again, the rsquared value is high because averaging across conditions decreases the amount of
variance. The regression analyses presented in the current paper will be computed only at
phonological phrase boundaries, both on a trial-by-trialbasis, and on the average
boundary production across-trials of the same condition, to allow comparison with
W&G's original results.
Although the LRB performed well, there are two significant problems with W&G's first
study, which the current study addresses. First, the LRB is a post-hoc model, tested on
the data it was designed to explain. Although W&G did conduct one subsequent test of
the LRB (Expt 2 in W&G 2004), this experiment involved only one syntactic
construction. Because of this, the current study was designed to evaluate the LRB's
predictions on a new set of syntactically varied materials. In addition, non-syntactic
factors, such as the interpretation of relative clauses, could be contributing to boundary
placement in the original set of items on which the formulation LRB was based. The
relative clauses contained in many of the items could be interpreted restrictively or nonrestrictively. W&G suggest that, in the absence of context, speakers followed the simpler
non-restrictive reading, since this reading does not require the instantiation of a contrast
set (as a restrictive relative clause does) (see Grodner, Gibson, & Watson, 2005). They
hypothesized that speakers tend to place boundaries before or after non-restrictive
32
relative clauses because of factors that go beyond Planning and Recovery. Specifically,
because non-restrictive relative clauses constitute new or aside information, speakers may
be more likely to set them off from the rest of the sentence in their own intonational
phrase. Indeed, in the third experiment of their 2004 paper, W&G found that when they
established contexts which disambiguated the restrictiveness of a relative clause with
prior context, speakers were more likely to produce boundaries before relative clauses
with a non-restrictive reading. Given this finding that the introduction of a new clause
can induce speakers to produce intonational boundaries for non-syntactic reasons, the
stimuli used for the current experiment did not contain relative clauses.
In addition to addressing problems with the original study, the current study was designed
to specify the predictions of the Recovery (LHS) component of the LRB. In earlier work,
Watson and Gibson considered other versions of the Recovery component, which took
into account more than the size of the most recently completed constituent, but settled on
the definition in (3) as a reasonable starting point. The current study was designed to
systematically compare three formulations of Recovery:
Incremental: A logical first proposal for how to measure left-hand distance is as a
measure of the distance back to the last boundary the speaker produced. We will refer to
this formulation as the Incremental version of recovery. Under this view, intonational
boundaries are a result of the physical needs of the speaker. The need to breathe or to
reset pitch induces the speaker to place boundaries at regular intervals such that the
likelihood of a boundary increases with the amount of material a speaker has produced
since the last boundary was produced. To test this alternative, we designed items that
included length manipulations at two different points in the sentence in order to see
whether the placement of a boundary earlier in the sentence influenced the probability of
boundary placement later in the sentence.
IntegrationDistance: Watson and Gibson (2001) first proposed the Recovery component
as a measure of integration distance of the upcoming constituent, or, the distance back to
the head with which the upcoming constituent was integrating. We will refer to this
formulation as the IntegrationDistanceversion of Recovery. Under this view, speakers
place boundaries before words that must be non-locally integrated because such
integration is more complex than local integration, as evidenced by work in self-paced
reading showing that readers slow down when reading words that must be integrated with
a non-local head compared to words that integrate with a local head (Gibson, 1998;
Grodner & Gibson, 2005; Gordon, Hendrick, & Johnson, 2001). The complexity of nonlocal integration as observed in reading could also be more complex in production, and
would induce speakers to produce boundaries at such points to recover from production
difficulty. We tested this alternative in two ways by designing items in which (1)
integration distance was varied while the size of the most recent completed constituent
was held constant, or (2) integration distance was held constant while the size of the most
recent constituent varied.
Semantic Grouping: The method of computing the Recovery weight that W&G used in
their original formulation of the LRB, in which Recovery is computed as the size of the
largest most recently completed constituent, will be referred to as the Semantic Grouping
version of Recovery. Under this view of boundary production, speakers produce words
and constituents that rely on one another for meaning in the same intonational group and
separate into different intonational words and constituents which don't depend on one
another. In order to test this view, we manipulated the length of the most recently
33
completed constituent to see if the presence of more semantically related material before
a possible boundary location led to a greater incidence of boundaries at that location.
Length manipulations designed to test the three versions of the Recovery component and
one version of the Planning component of the LRB were obtained by varying the length
of three post-verbal constituents in sentences that took the form: Subject Verb Direct
Object Indirect Object modifier. A post-verbal direct object was either short (the
chapter) or long (the chapter on local history); an indirect object was either short (to the
students) or long (to the student ofsocial science). Finally, a modifier was either short
(yesterday) or long (after thefirst midterm exam). These three independent
manipulations resulted in eight conditions. An example item is presented in (5).
(5)
a. Short direct object, Short indirect object, Short modifier
The professor assigned the chapter to the students yesterday.
b. Long direct object, Short indirect object, Short modifier
The professor assigned the chapter on local history to the students yesterday.
c. Short direct object, Long indirect object, Short modifier
The professor assigned the chapter to the students of social science yesterday.
d. Long direct object, Long indirect object, Short modifier
The professor assigned the chapter on local history to the students of social science
yesterday.
e. Short direct object, Short indirect object, Long modifier
The professor assigned the chapter to the students after the first midterm exam.
f. Long direct object, Short indirect object, Long modifier
The professor assigned the chapter on local history to the students after the first midterm
exam.
g. Short direct object, Long indirect object, Long modifier
The professor assigned the chapter to the students of social science after the first midterm
exam.
h. Long direct object, Long indirect object, Long modifier
The professor assigned the chapter on local history to the students of social science after
the first midterm exam.
Predictions
To investigate the three possible quantifications of Recovery, and to evaluate the overall
accuracy of the LRB, we compared the LRB's predictions to actual speakers' boundary
placement at four phonological phrase boundaries in each sentence. These points,
indicated in (6), include: (1) between the subject and the verb, (2) between the verb and
the direct object, (3) between the direct object and the indirect object, and (4) between the
indirect object and the modifier.
(6) The teacher b1
assigned 12 the chapter (on local history) 13to the students (of social
science) 14 yesterday / after the first midterm exam.
34
The LRB does not predict any differences in boundary occurrence at the first two
sentence positions. Between the subject and the verb, the total LRB weights are
comparable because the size of all three versions of the Recovery component (i.e. the
length of the subject) is the same in every condition, and the Planning component (i.e. the
entire verb phrase) in all conditions is always relatively large (2 4 phonological phrases).
Therefore, the LRB predicts minimal differences across conditions. Between the verb
and the direct object, the LRB weights are the same for the following reasons: First, the
size of all three versions of the Recovery component is the same in all conditions (i.e. two
phonological phrases corresponding to the subject and the verb); second, in every
condition of every item, the direct object is an obligatory argument of the preceding verb,
so the Planning component will have a weight of zero because it disallows boundaries
between heads and their obligatory arguments.
The three different versions of Recovery make different predictions at both the third and
fourth points in the test sentences, and we will elaborate each in turn.
Incremental
In order to test the predictions of an incremental version of Recovery, we included two
length manipulations within each sentence. In this way, we could see whether the
presence (vs. absence) of a boundary at a location early in the sentence would lead to
fewer boundaries at the later location, irrespective of the size of the adjacent constituents
at the second location. For example, we hypothesized that if speakers did not place
boundaries after a short direct object, as in (5a), they might be more likely to place a
boundary after a short indirect object. Although the Semantic Grouping version of
Recovery would predict few boundaries between the indirect object and modifier in this
example, an Incremental version would predict more simply because the speaker has
produced at least two phonological phrases since his/her last boundary.
Integration Distance
The stimuli were created to allow for two critical comparisons to investigate the
Integration Distance version of Recovery. The first is between sentences (5b) and (5c) or
(5f) and (5g). In (5b), at "yesterday," the size of the most recently completed constituent
is one phonological phrase ("to the students") whereas the integration distance back to
"assigned"-which "yesterday" must be integrated with-is three phonological phrases
("the chapter / on local history / to the students"). Conversely, in sentence (5c), while the
size of the most recently completed constituent has increased to two phonological phrases
("to the students / of social science"), the integration distance back to the verb is still
three phonological phrases ("the chapter / to the students / of social science"). Whereas
Semantic Grouping predicts more boundaries before the modifying adverb "yesterday" in
condition (5c) than in (5b), and in (5g) than in (5f), Integration Distance predicts an equal
boundary probability between the two conditions.
The predictions of Integration Distance also differ from those of Semantic Grouping at
the phonological phrase boundary between the indirect object and the modifier when the
direct object has been either long or short but the indirect object remains constant.
Specifically, when the indirect object length is matched, Integration Distance predicts
more boundaries when the direct object is long than when the direct object is short,
because when the direct object is long, the modifier has a longer distance to integrate
back to the verb than when the direct object is short. In contrast, Semantic Grouping
35
predicts an equal probability of boundary production in either case, as the most recently
completed constituent (the indirect object) is the same length.
Semantic Grouping
The Semantic Grouping version of Recovery predicts main effects of length at the two
testing points. Between the direct object and the indirect object-"the chapter (on local
history)" / "to the students (of social science)"-Semantic Grouping predicts a main
effect of direct object length such that speakers will place more boundaries after long
direct objects than short direct objects. Semantic Groupingalso predicts a main effect of
indirect object length such that speakers will place more boundaries before long indirect
objects than short indirect objects. Between the indirect object and modifier, Semantic
Groupingpredicts main effects of indirect object length, such that speakers will place
more boundaries after long indirect objects than short indirect objects. It also predicts a
main effect of modifier length such that speakers will place more boundaries before long
modifiers than short modifiers.
Method
Particpiants
Forty-eight native English speakers from the MIT community participated in the study
for $10.00 each. Participants were run in pairs, and each member of the pair was
randomly assigned to either the role of Speaker or the role of Listener. Data from three
of the twenty-four pairs of subjects could not be used due to poor recording quality.
Productions from eighteen of the remaining twenty-one pairs were coded for intonational
boundaries by two blind coders. Productions from all twenty-one successfully recorded
subjects were analyzed for word duration and silence data.
Materials and Design
Length of direct object, length of indirect object, and length of modifier were
manipulated in a 2x2x2 design to create thirty-two stimulus sets like those in (5).
The long direct object and long indirect object conditions were created by adding
a modifier phrase (or non-obligatory argument phrase) to the object noun phrase (e.g. the
bouquet of thirty roses, the turkeys with homemade stuffing, the chapter on local history).
All direct objects and indirect objects in the short condition had three syllables, while the
long conditions had seven or eight syllables. The short modifiers were temporal
modifiers (in 23 items) or adverbs (nine items) comprised of one or two words (two to
four syllables), but always only one phonological phrase (e.g. secretly, last night, on
Sunday). The long modifiers were always temporal modifiers containing five words,
which were comprised of 3-4 phonological phrases.
Twelve of the 32 experimental items contained an information structure
manipulation in the long modifier condition. Specifically, the long modifier in these
items contained a new. clause (e.g. afterfilming had already begun, before the crime was
committed). In order to discover whether speakers place more boundaries before the
introduction of a new clause, we computed analyses of variance on two subsets of items:
those with new clauses in the modifier and those without. The analyses were virtually
identical to the analyses presented below, which were computed across all thirty-two
items. Thus, the presence/absence of a clause boundary does not seem to have affected
36
the probability of the productions of an intonational boundary. Any divergence across
the three analyses will'be noted.
The materials were presented in a Latin Square design, resulting in eight lists. Each
participant saw only one of the lists, which was presented in a random order.
Experimental items were randomly interspersed with 44 fillers, which were comprised of
items from two other unrelated experiments, with different syntactic structures. A full set
of experimental items can be found in the Appendix.
Procedure
The experiment was conducted using Linger, a software platform for language processing
experiments. 3 Two participants-a speaker and a listener-were included in each trial,
and sat at computers in the same room such that neither could see the other's screen. The
speakers were instructed that they would be producing sentences for their partners (the
listeners), and that the listeners would be required to answer a comprehension question
about each sentence immediately after it was produced. Each trial began with the speaker
being presented with a sentence on the computer screen to read silently until s/he
understood it. The speaker then answered a multiple-choice content question about the
sentence, to ensure understanding. If the speaker answered correctly, s/he proceeded to
produce the sentence out loud once. If the speaker answered incorrectly, s/he was given
another chance to read the sentence, and to answer a different question about it. The
speaker always produced the sentence after the second question whether or not s/he got
the second question right.
The listener sat at another computer, and saw a blank screen while the speaker went
through the procedure described above for each sentence. After the speaker produced a
sentence out loud for the listener, the listener would press the space bar on his/her
computer, whereupon s/he was presented with a multiple-choice question about the
content of the sentence that was just produced. Listeners were provided feedback when
they answered a question incorrectly.
Trials where one or both of two blind coders identified a disfluency in the production
were excluded from analysis, following the method of W&G, accounting for 4.6% of the
data. Trials where either a) the speaker answered both comprehension questions
incorrectly or b) the listener answered his/her comprehension question incorrectly
accounted for 3.2% of the data. We conducted all analyses reported below on a) all trials
and b) only trials without any incorrect responses, and the results of both analyses were
not different. Therefore, we report the results from all fluent trials below.
Boundary identification
Each sentence was recorded digitally, and analyzed using the PRAAT program (Boersma
& Weenink, 2006). Each production was coded by two expert coders (neither of whom
was an author) for intonational boundaries using a subset of the ToBI intonational
labeling conventions (Silverman et al., 1992). Both coders were blind to the predictions
of the experiment. The strength of a boundary was marked by each of the coders using
the following standard break indices and disfluency notation: 4 - full intonational phrase
boundary (IP); 3 - intermediate phrase boundary (ip); 0, 1, 2 - no phrase boundary; P hesitation pause; D - disfluency. Most of the non-phrasal boundaries were coded as "1".
We therefore collapsed all of 0, 1, and 2 as the category 1. The raw numerical labels (i.e.
3Linger was written by Doug Rohde, and can be downloaded at: http://tedlab.mit.edu/~dr/Linger/
37
1, 3, 4) were grouped in two different ways for two separate analyses reported below.
Trials which contained hesitations or disfluencies were excluded from analysis.
Boundary identification is not a straightforward, objective task. Although several
acoustic measures, such as silence and lengthening, have been found to correlate with
raters' identification of intonational boundaries (see Wightman, et al., 1992), even expert
coders are not in perfect agreement about the presence of boundaries (Pitrelli, et al.,
1994; Dilley, et al, 2006; Syrdal & McGory, 2000 Yoon, et al., 2004). We therefore
computed the inter-rater reliability of the two coders. Furthermore, we computed the
correlation of the coders' identification of boundaries with both the duration of the preboundary word and any silence that accompanied the boundary location.
Each trial was annotated for word boundaries and silence by one of three coders, none of
whom were authors on the current paper, or ToBI coders for the present study. These
coders were blind to the hypotheses of the study. Using output from the PRAAT
program, the coders annotated all word boundaries, as well as any perceptual silence that
was distinct from silence due to stop closure.
Inter-rater reliability
Reliability between ToBI coders was measured by calculating the proportion of the
instances in which the two transcribers agreed on the label of a word boundary over the
number of possible agreements, as described in Pitrelli et al. (1994). To avoid artificial
inflation of the agreement numbers, we excluded the final word in each sentence from
analysis, as these words, being utterance final, always received an obligatory break index
label of '4' from both coders, and therefore contributed a trivial point of agreement to the
analyses. As such, we computed, for example, nine points of agreement for a sentence
composed of ten words.
We computed agreement between the coders in two different ways. First, we compared
the coders' raw break indices (BIs) (1, 3, 4), resulting in a total agreement of 78.8%.
Second, we computed agreement when we compared ips (BI of 3) or IPs (BI of 4)
boundaries to the absence of a boundary (BI of 1), which resulted in an overall agreement
measure of 82.0% between the two coders. Finally, we computed agreement when both
raters indicated an IP (BI of '4'). This calculation resulted in overall agreement of
94.9%.
All of the above agreement numbers are consistent with previous measures of boundary
agreement in ToBI (e.g. Pitrelli, et al., 1994; Dilley, et al., 2006). However, in order to
effectively use the data we collected from two coders, the two measures of boundary
proportion that we will present in the results section come from the average of the two
coders labels. We will present one set of analyses which we conducted on the average of
the coder's boundary decisions where only IPs (i.e. Break indices of '4') are considered
to be boundaries. The second set of analyses is based on the coders' data where IPs and
ips (i.e. Break indices of '4' or '3') are considered to be boundaries. In both cases, the
binary distinction of each individual coder (of boundary vs. not boundary) was expanded
into a ternary distinction where a phonological phrase could be (a) 0: a non-boundary as
indicated by both coders, (b) .5: a boundary for one of the coders and not the other, or (c)
1: a boundary indicated by both coders.
Acoustic correlates of boundaries
To further test the accuracy of the coders' labels of intonational boundaries, we correlated
several acoustic measures with the coders' ToBI labels. We gathered measures of (a) the
duration of each word that preceded a phonological phrase boundary, (b) the duration of
any silence that followed a phonological phrase boundary, and (c) the sum of each of
these two measures. We then correlated each of these three acoustic variables with the
two formulations of boundary labels described above; first, where only IPs are considered
to be boundaries; second, where IPs and ips are considered to be boundaries.
When we considered only IPs as boundaries, we observed a correlation between
pre-boundary word duration and average coder boundary label (r, =. 091, N = 2636,
p<. 001), indicating that boundaries were more likely to be identified after lengthened
words. We also observed a correlation between post-boundary silence and boundary
label (r' = .318, N = 2636, p<. 001), such that measurable silence was more likely to
occur when coders indicated a boundary. Finally, we observed a correlation between the
coders' average boundary label and the sum of the word duration and post-word silence
(?=.234, N = 2636,p<.001). Because the correlation between the presence of silence
and the labeling of an IP was the highest of these three analyses, we used the presence of
silence as the dependent measure in a series of ANOVAs conducted at each critical
constituent.
When we considered IPs or ips to be boundaries, we observed a correlation
between pre-boundary word duration and average coder boundary label (r 2 =. 153, N =
2625, p<. 001), indicating that boundaries were more likely to be identified after
lengthened words. This correlation suggests that word duration may be a better cue for
intermediate boundaries than for full intonational boundaries as this correlation is better
than the one performed on IPs above. We also observed a correlation between postboundary silence and boundary label (r2 = .097, N = 2625, p<. 001), such that measurable
silence was more likely to occur when coders indicated a boundary. This correlation
suggests that silence is a stronger cue to full intonational boundaries (IPs) than to
intermediate boundaries (ips) as this correlation is lower than the one performed on IPs
above. Finally, we observed a correlation between the coders' average boundary label
and the sum of the word duration and post-word silence (r2 =.205, N = 2625, p<. 001).
Because the combination of word duration and post-word silence proved to correlate
better with coders' labels of '3' or '4' than either cue alone, we used this measure in a
different series of ANOVAs'conducted at each critical constituent boundary.
The above correlations are noteworthy for the following reason: Because the
experiment was conducted using a fully between-subject Latin-square design, each
subject produced only one token of each word used in this analysis. Moreover, though
the actual words being compared were all two-syllable words with initial stress, they
differed greatly in terms of segmental characteristics. Individual speaker and word
variation add a large amount of noise to this analysis. However, despite the variability,
we still observed a strong relationship between boundary perception and acoustic
measures. The fact that the acoustic measures correlated with the perceived boundaries
suggests that the coders may have been using these and other acoustic cues to code
boundary locations.
Data analysis
We will present the results of the experiment in several different ways to investigate the
effectiveness of the three versions of the Recovery and the one version of Planning which
we have proposed to account for boundary placement. First, we present the results of a
series of ANOVAs, where the dependent measure corresponds to four different
quantifications of boundaries, to compare the specific predictions of the models at the
four critical points indicated in (6). Second, we present results of a series of regression
models which test the success of predictions across all the phonological phrase
boundaries in each sentence, as indicated in (7).
(7) The teacher 1iassigned 12 the chapter 13 on local history |4 to the students 15 of social
science 16 yesterday.
We will first present the results of analyses where 3's or 4's are considered to be
boundaries, following W&G. Using these criteria, boundaries were indicated by both
coders 33.1% of the time, and by one of the two coders 48.1% of the time. Along with
these analyses, we will present the results of ANOVAs conducted with the sum of word
duration and following silence as the dependent measure, as this acoustic measure
correlated most highly with combined '3' and '4' labels.
We will next present the results of the analyses where only 4's are considered to be
boundaries. Using these criteria, boundaries were indicated by both coders 6.1% of the
time, and by one of the two coders 14.2% of the time. Along with these analyses, we will
present the results of analyses conducted on post-boundary silence, as this measure
correlated most highly with coders' labels of '4.'
In addition to the analyses of variance performed at each critical phonological phrase
boundary in the sentences, we also performed a series of regression analyses, to see
which version of the LRB model accounted for the most variance in the speakers'
productions. We will present regression analyses performed on the means of boundary
productions at each phonological phrase boundary across all of the sentences, and
analyses performed on a trial-by-trialbasis.
It should be noted that the regressions testing the Incremental version of recovery differ
from those testing Semantic Grouping and Integration in three important ways.
First, regressions using the latter two versions of Recovery can easily be performed on
the sentence means, as the piedictions of these models do not change across trials; In
contrast, the predictions of the Incremental version do change across trials as they are, by
definition, contingent upon what has previously happened in the sentence. In order to
perform regression analyses on means in this case, we computed the average distance (in
phonological phrases) back to the location of the last boundary using the following
method. Consider (8), where P, ... Pn indicate phonological phrases and bi indicates the
proportion of boundaries between Pi and Pi,1:
(8) P1
I P2 IP3 I P 4
b,
b2
...
b3 ...
At the first possible boundary location (between P, and P2 ), the average distance back to
the last boundary is always one phonological phrase (i.e. the distance back to the
beginning of the sentence). At the second possible boundary location (between P2 and
P3 ), the average distance back is [1 + (1-b1 )]. At the third candidate boundary location
(between P3 and P 4), the average distance back is 1 + (1-b2)+ (1- (1- b2 ) * (1- bi)). The
proportion of boundaries at later locations is computed similarly.
Second, analyses on the Semantic Grouping and Integration Distance versions of
Recovery can easily be performed on the average of the two coders' data. However,
because the incremental predictions are contingent on the labels of a particular coder,
they cannot be averaged. Therefore, we performed regressions on the two coders' data
individually in order to test the Incremental version of Recovery.
Finally, it is not possible to use the Incremental version of Recovery to predict the
acoustic measures of word duration and silence. The Incremental version of Recovery's
predictions about where boundaries will occur is based on the location of previous
boundaries. Without a dichotomous way of using duration and silence to define the
location of boundaries, there is no way of assessing where the previous boundary has
been placed. Moreover, even if there were a way of defining the location of the last
boundary, it is not clear how increased distance from that boundary would translate into
the continuous variables of word duration and silence.
Results
Boundaries as Intermediate or Full Intonational Phrases: ANOVAs
We performed a series of 3 x 2 analyses of variance with Participants or Items as the
random factor at the sentence positions indicated in (6), which were a) after the Subject
NP, b) after the Verb, c) after the direct object, and d) after the indirect object. The
dependent variables were the averaged ratings of the two coders when IPs or ips were
considered boundaries and the corresponding acoustic measure of the duration of the preboundary word and following silence. The independent variables in each analysis were
a) the length of the direct object (short, long), b) the length of the indirect object (short,
long), and c) the length of the modifier (short, long). In all cases, analyses were
conducted only on fluent trials.
Subject NP:
A 3 x 2 ANOVA with boundary percentage after the Subject NP as the dependent
measure revealed no effect of direct object length on boundary placement, F 1(1,17) < 1,
F2(1,31)=1.752, p=.195. There was, however, a marginal main effect of length,
F1(1,17)=4.99, p<.05, F2(1,31)=3.39, p=.075, such that more boundaries occurred after
the subject NP when the upcoming indirect object was long than when it was short.
There was no effect of modifier length, Fl<1; F2(1,31) = 2.17, p=.15. There were no
effects of direct object length, indirect object length, or modifier length on the duration of
the phrase-final word and following silence (4 of 6 F's < 1; all p's > .14).
Verb:
A 3 x 2 ANOVA revealedno effects of direct object length, indirect object length, or
modifier length on the on the probability of boundary placement after the verb, as
quantified by ToBI label (4 of 6 F's <1; all p's > .16), or by the sum of phrase-final word
duration and silence (4 of 6 F's <1; all p's >.24)
1
IL
Long DO
t0.4 C
200.2W
0
Short 10 /Short
Mod
Short 10/ Long
Mod
Long 10/ Short
Mod
Long 10 /Long
Mod
Figure 1: Proportionof boundariesplaced after the Direct Object NP in all conditions.
Direct object:
The percentages of boundary placement between the direct object and indirect object, as
determined by coders' ToBI labels, are presented in Figure 1. There was a main effect of
direct object length such that boundaries occurred more often after long direct objects
than short direct objects F1(1,17)=76.17, p<.001, F2(1,31)=132.95, p<.001. An effect of
indirect object length such that speakers placed more boundaries before long indirect
objects than short indirect objects approached significance in the participants' analysis
(F1(1,17)=2.76, p =. 11), but not in the items' analysis, F2(1,31) < 1 . There was also no
effect of modifier length in this position, F's<1.
In the acoustic analyses, we also observed a main effect of direct object length such that
the final word of a long direct object was longer and followed by longer silence (61ms)
than a short direct object (55ms), F1(1,20)=14.35, p<.001, F2(1,31)=4.33, p<.05. There
were no effects of indirect object length or modifier length at this position (All F's < 2;
all p's > .20).
1
C 0.8
06
E Short DO
0.4
U Long DO
C
W 0.2
0.
Short 10/ Short Mod Long 10/ Short Mod Short 10/ Long Mod Long 10/ Long Mod
Figure 2: Proportionof boundariesplacedafter the indirect object NP in all conditions.
42
Indirect object:
The percentages of boundary placement between the indirect object and the modifier are
presented in Figure 2. There was a main effect of indirect object length such that more
boundaries occurred following long indirect objects than short indirect objects, F 1(1,17)
= 63.09, p<.001; F2(1,31) = 38.49, p<.001. There was also a main effect of modifier
length such that more boundaries preceded long modifiers than short modifiers, F1(1,17)
=
48.23, p<.001; F2(1,31) = 34.74, p<.001.
In the acoustic analyses, we observed a main effect of indirect object length such
that the combined duration of the final word of a long indirect object and any following
silence (56ms) was longer than that of a short indirect object (51ms), F1(1,20)=14.35,
p<.OOl, F2(1,31)=4.33, p<.05. We also observed a main effect of modifier length such
that the combined duration of the final word of the indirect object and any following
silence was longer when the modifier was long (56ms) than when the modifier was short
(5ims), F1(1,20) =2.82, p<.05, F2(1,31)
=
7.29, p<.05.
We performed two critical comparisons at the indirect object/modifier boundary to test
the predictions of the Integration version of Recovery. First, we compared the proportion
of intonational boundaries between the indirect object and the modifier when the distance
back to the verb (i.e. integration distance) was matched. This condition was satisfied in
cases where the direct object was long and the indirect object was short, or when the
direct object was short and the indirect object was long. An independent samples t-test
comparing these two conditions was conducted on the coders' data revealed a significant
effect of indirect object length on the boundary labels t(259) = -5.94, p<.001, and on the
acoustic correlates (word duration and silence), t(259) = -3.06, p<.005), such that
speakers placed more boundaries between indirect object and modifier when the indirect
object was long, regardless of the integration distance back to the verb. This result is in
contrast to the predictions of the Integration version of Recovery.
The second test of Integration Distance concerned cases where the size of the most
recently completed constituent stayed the same, while the integration distance back to the
verb varied. This condition is satisfied in when the indirect object is the same length, but
the direct object is varied. Independent samples t-tests conducted on the coders' ratings
and the acoustic measures revealed no difference between boundary placement after a
long indirect object regardless of the length of the direct object (all p's>.4), indicating
that the Integration Distance version of Recovery did not account for boundary placement
after the indirect object phrase.
In accordance with the predictions of the Incremental version of Recovery, we observed a
significant interaction between direct object length and indirect object length such that
more boundaries preceded the modifier in the short direct object / short indirect object
condition than in the long direct object/short indirect object condition F1(1,17)=6.40,
p<.05; F2(1,31)=9.63, p<.005. In addition, we observed a significant 3-way interaction
between direct object length, indirect object length and modifier length (F1=5.46, p<.05;
F2=4.31, p<.05), such that in cases where the modifier was long, speakers
overwhelmingly placed a boundary before it, regardless of what had preceded that point.
This interaction only reached significance in the participants' analysis of items that did
not introduce new clauses in the modifier phrase, F1 (1,17) =5.57, p<.05; F2 (1,19)
=2.41, p=.14).
Boundaries as Full IntonationalPhrases: ANOVA
We performed a series of 3 x 2 analyses of variance with Participants or Items as the
random factor at the sentence positions indicated in (6), which were a) after the Subject
NP, b) after the Verb, c) after the direct object, and d) after the indirect object. The
dependent variables were the averaged ratings of the two coders when only IPs were
considered boundaries, and the corresponding acoustic measure of the duration of postword silence. The independent variables in each analysis were a) the length of the direct
object (short, long), b) the length of the indirect object (short, long), and c) the length of
the modifier (short, long). In all cases, analyses were conducted only on fluent trials.
Subject NP
A 3 x 2 ANOVA with boundary as the dependent measure revealed no effect of direct
object length (F's < 1), no effect of indirect object length (F's < 1), and no effect of MOD
length (F1 (1, 31)=2.09, p = .17; F2 (1,17)= 1.88, p = .18). A 3 x 2 ANOVA with
duration of silence as the dependent measure revealed no effect of direct object length
(F's <1), and no effect of indirect object (F1<1, F2=1.24, p=.27), but a marginal effect of
modifier length, such that longer silence followed the Subject NP when the sentence
contained a long modifier (20ms) than when it contained a short modifier (lOims),
F1(1,20) =4.66, p<.05 F2(1, 31) = 3.324, p = .08.
Verb
A 3 x 2 ANOVA revealed no effects of direct object length, indirect object length, or
modifier length on the on the probability of boundary placement after the verb, as
quantified by ToBI label (All F's <2; all p's > .25), or by post-word silence (4 of 6 F's <
2; all p's > .12).
Direct Object
The proportion of boundaries that occurred between the direct object and the indirect
object are plotted in Figure 3. A 3 x 2 ANOVA with boundary as the dependent measure
revealed a main effect of direct object length (F1(1,17) = 34.00, p <.001; F2(1,31) =
56.60, p <.001), such that more boundaries occurred after long direct objects (24.6%)
than short direct objects (7.2%). The analysis also revealed a main effect of indirect
object length (F1(1,17) = 11.42, p <.005; F2(1,31) = 4.92, p < .05), such that speakers
placed more boundaries before long indirect objects (18.6%) that before short indirect
objects (13.2%). Finally, there was a suggestion of an effect of modifier in the
participants analysis, in the opposite of the predicted direction (F 1(1,17) = 5.09, p < .05),
but this effect was not reliable in the items analysis (F2(1,31)= 1.69, p = .2 1).
A 3 x 2 ANOVA with duration of silence as the dependent measure revealed a
main effect of direct object length (F1=6.75, p<.05; F2=4.30, p<.05) such that longer
silences followed long direct objects (45ms) than short direct objects (21ms). There was
no effect of indirect object length (F1=2.05, p=.17; F2=1.88, p=.18), and no effect of
modifier length, F's<1.
44
1
C
0
a
0.8
0.
0-
0.6
. 0.4-
D
E Short DO
Long DO
0.2
0
Short 10 / Short Short 10/ Long Long 10/ Short
Mod
Mod
Mod
Long 10/ Long
Mod
Figure 3: Proportionof boundaries(as determined by average of coder's '4' labels)
between the direct object and the indirectobject in all conditions.
Indirect Object
The proportion of boundaries that occurred between the indirect object and modifier are
plotted in Figure 4. A 3 x 2 ANOVA with boundary as the dependent measure revealed
no main effect of direct object length (F's < 1). The analysis did reveal a main effect of
indirect object length (F1(1,17) = 5.75, p<.05; F2(1,31) = 12.00, p < .005), such that
speakers placed more boundaries after long indirect objects (22.6%) than before short
indirect objects (14.0%). Finally, this analysis revealed a main effect of modifier length
(FI(1,17) = 39.42, p < .001; F2(1,31) = 39.23, p < .001), such that speakers placed more
boundaries before long modifiers (28.9%) than before short modifiers (8.3%).
A 3 x 2 ANOVA with duration of silence as the dependent measure revealed no effect of
direct object (F's<1), and no effect of indirect object (F1=2.81, p=.1 1; F2=1.63, p=.21);
however, this analysis did reveal a main effect of modifier length (F 1=15.95, p<.005;
F2=12.83, p<.005), such that longer silences occurred before long modifiers (42ms) than
before short modifiers (12ms).
0.8
e
0
0.6
0.4
NShort DO
E Long DO
0.2
0
0
0
Short 10/ Short Short 10/ Long Long 10 / Short
Mod
Mod
Mod
Long 10/ Long
Mod
Figure 4: Proportionof boundaries (as determined by average of coder's '4' labels)
between the indirect object and the modifier in all conditions.
45
We performed two critical comparisons at the indirect object/modifier boundary to test
the predictions of the Integration version of Recovery. First, we compared the proportion
of intonational boundaries between the indirect object and the modifier when the distance
back to the verb (i.e. integration distance) was matched. This condition was satisfied in
cases where the direct object was long and the indirect object was short, or when the
direct object was short and the indirect object was long. An independent samples t-test
comparing these two conditions was conducted on both the coders' boundary data. This
test revealed a significant effect of indirect object length (t(260) = -2.32, p<.05) such that
speakers placed more boundaries between the indirect object and the modifier when the
indirect object was long (22.9%), than when the indirect object was short (13.7%),
regardless of the integration distance back to the verb. This result is in contrast to the
predictions of the Integration version of Recovery. In the analysis of post-word silence,
this effect was not significant, t(308) = -1.20, p=.23.
The second test of Integration Distance concerned cases where the size of the most
recently completed constituent stayed the same, while the integration distance back to the
verb varied. This condition is satisfied in cases where the indirect object is the same
length, but the direct object is varied, a condition which was satisfied in two places in this
current experiment. When the indirect object is long, speakers place boundaries after the
indirect object 22.9% of the time when the direct object is short and 22.4% of the time
when the direct object is long. When the indirect object is short, speakers place
boundaries after the indirect object 14.2% of the time when the direct object is short and
13.7% of the time when the direct object is long. Two independent samples t-tests
conducted on the coders' boundary ratings revealed no difference between boundary
placement after a long indirect object regardless of the length of the direct object (ti(261)
=.128, p =.898; t2 (263) = .125, p = .900), indicating that speakers produced comparable
numbers of boundaries when the size of the largest recently-completed constituent was
the same, regardless of the Integration Distance back to the verb. The pattern of results
was the same for the analysis of post-word silence, (ti(306) = .024, p = .980; t2(312)=
.125, p =.717). Once again, these results support the Semantic Grouping version of
Recovery over the Integration version of Recovery.
Finally, the Incremental version of Recovery predicted a two-way interaction between the
length of the direct object and the length of the indirect object, as pictured in Figure 1,
such that speakers would place more boundaries after a short direct object and a short
indirect object than after a long direct object a short indirect object. However, this
interaction did not approach significance in the analysis of boundaries (F's <1) or in the
analysis of post-word silence (F's<1), failing to show support for the Incremental version
of Recovery.
Boundaries as Intermediate or Full Intonational Phrases: Regressions
A series of multiple regression analyses was conducted to predict the average coders'
label of a '3' or '4' from the three versions of Recovery-Incremental (2 versions),
Integration Distance, Semantic Grouping-and the single version of Planning. The
regressions were performed on a trial-by-trial basis, and on the condition means. A
summary of the regression models tested is presented in Table 1.
As is evident from the table, every model tested accounted for a significant
amount of the variance in the speakers' production of boundaries. However, the best
model in both groupings of the data is the one using the Semantic Grouping version of
Recovery. For example, when tested on all of the production data, trial-by-trial, this
46
model accounted for more variance than any of the other three models, R2 = .44,
F(2,2325) = 898.86, p <.001. In addition, in both models which include the Semantic
Grouping version of Recovery, both the left and the right factors accounted for a
significant amount of the variance, indicating that an increase in a) the size of the largest
recently-completed left-hand constituent or b) the right-hand constituent increase the
probability that a speaker will produce a boundary at a candidate phonological phrase
boundary. The model using the Integration version of Recovery, tested on all of the data,
trial-by-trial, also accounts for a significant amount of the variance in speakers' boundary
production, R2 =.37, F(2,2325) = 685.04, p <.001). It predicts that an increase in a) the
distance back to the head with which an upcoming constituent must be integrated or b)
the size of the upcoming constituent will increase the probability of a boundary at a
candidate phonological phrase boundary. Although significant, this model does not
account for as much variance as the Semantic Grouping version. Finally, the models
using the Incremental version of Recovery, based on each of the speakers' data, both
account for a significant amount of the variance in the individual coders' production data.
However, these models are less successful than the model using the Semantic Grouping
version of Recovery. In several cases, the Incremental factor is negatively correlated
with boundary production, indicating that an increase in the distance back to the last
boundary leads to a decrease in the probability that a speaker will produce a boundary at
a candidate phonological phrase boundary.
Beta SE
.
Coder 2 Incremental Left 0.08 0.01
0.06 .
Planning Right
Coder 1 Incremental Left 0.03 0.01
F and p-value
0.13 0.01
0.001
Integration Left
0.20 0.01
0.001 0.37 F(2,2325) = 685.04, p<.001
Ri ht
0.04 0.05
0.001
Semantic Grouping Left
0.38 0.01
0.001 0.44 F(2,2325)= 898.86, p<.001
0.02 0.00
0.001
Planning Right
Coder 2 Incremental Left 0.22 0.03
Planning Right
0.09 0.02
Coder 1 Incremental Left 0.32 0.09
Planning Right
-j Integration Left
'8
R
0.001 0.07 F(2,2490)= 90.88, p<.001
0.01
0.017 0.14 F(2,2330) = 196.74, p<.001
Planning Right
Plannin
e
p-value
Planning Right
U Semantic Grouping Left
Planning Right
0.001 0.66 F(2,37) = 35.54, p<.001
0.001
0.001 0.51 F(2,37) = 18.91, p<.001
0.16 0.03
0.001
0.19 0.02
0.001 0.77 F(2,37) = 62.21, p<.001
0.04 0.02
0.008
0.38 0.02
0.001 0.91 F(2,37)= 93.43,p<.001
0.02 0.01
0.041
Table 1: Summary ofregression models tested on data where 3's or 4's are boundaries.
Models were tested on allfluent trials on a trial-by-trialbasis, and on the means of each
condition.
Boundaries as Full Intonational Phrases: Regression
A series of multiple regression analyses was also conducted to predict the average
coders' label of a '4' from the three versions of Recovery-Incremental (2 versions),
Integration Distance, Semantic Grouping-and the single version of Planning. The
47
regressions were performed on all fluent trials, on a trial-by-trial basis, or on the
condition means. A summary of the regression models tested is presented in Table 2.
As is evident from the Table, the results of these regressions pattern with those
conducted on the data when boundaries were considered to be labels of '3' or '4.' The
models that use the Semantic Grouping version of Recovery account for as much or more
variance as the other versions of Recovery.
|Beta ISE p-value 1R2
-
Coder 2 Incremental Left 0.02 0.00
F and p-value
0.001 0.02 F(2,2526) = 24.44, p<.001
Planning Right
0.001
0.02 0.00
Coder 1 Incremental Left 0.00 0.01
A Planning Right
0.05 0.01
0.333 0.05 F(2,2526) = 59.43, p<.001
0.001
Integration Left
0.06 0.01
0.001 0.09 F(2,2330)= 111.42, p<.001
Planning Right
0.02 0.00
0.001
Semantic Grouping Left
0.11 0.01
0.001 0.10 F(2,2330) = 123.25, p<.001
Planning Right
0.01 0.00
0.001
Coder 2 Incremental Left 0.04 0.01
2 Planning Right
0.02 0.01
0.001 0.62 F(2,37) = 30.39, p<.001
0.001
Coder 1 Incremental Left 0.07 0.09
e Planning Right
0.07 0.01
0.001 0.68 F(2,37) = 39.37, p<.001
0.001
-
Integration Left
0.06 0.01
0.001 0.61 F(2,37)= 28.30, p<.001
O
Planning Right
Semantic Grouping Left
0.02 0.01
0.10 0.01
0.005
0.001 0.65 F(2,37)= 34.86, p<.001
Planning Right
0.02
0.01
0.03
Table 2: A summary of regressionmodels tested on data where 4's are boundaries.
Models were tested on allfluent trials on a trial-by-trialbasis, and on the means of each
condition.
Boundaries as Word Duration and Silence: Regression
A series of multiple regression analyses was also conducted to predict the length of the
final word of a phonological phrase, and following silence, from either the Integration
Distance or Semantic Grouping version of Recovery, and the single version of Planning.
The regressions were performed on a trial-by-trial basis, or on the condition means, and
were performed on all of the data, as described in the Method section. A summary of the
regression models tested is presented in Table 3.
Similar to the regression analyses reported above, the regressions that use the
Semantic Grouping version of Recovery perform better than those which use the
Integration version of Recovery.
48
Table 3: A summary of regressionmodels tested on the data comparingthe predictionsof
Beta SE
-
p-value R2
F and p-value
0.09 F(2,2491) = 123.86, p<.001
Integration Left
0.03 0.00
0.001
Planning Right
0.03 0.02
0.001
Semantic Grouping Left
0.08 0.01
0.001
-Planning Right
0.02 0.00
0.001
Integration Left
0.02 0.01
0.001
Planning Right
Semantic Grouping Left
0.02 0.00
0.06 0.01
0.001
0.001 0.85 F(2,37) = 105.03,p<.001
Planning Right
0.01
0.001
0.00
0.13 F(2,2491) = 182.74, p<.001
0.58 F(2,37) = 25.64, p<.001
1_j
each model to the duration ofthe final word of each phonologicalphrase and any
following silence. Models were tested on all data on a trial-by-trialbasis, and on the
means of each condition.
Discussion
The aim of the current study was to investigate the relationship between syntactic
structure and intonational phrasing, and, specifically, to test the predictions of the LeftRight Boundary Hypothesis (LRB), proposed by Watson & Gibson (2004). The LRB
predicts the probability of a boundary according to the size of the syntactic material that
precedes a phonological phrase boundary (Recovery) and the size of the material that
follows a phonological phrase boundary (Planning). We evaluated the claims of the LRB
and some related hypotheses by designing materials which allowed us to compare three
different versions of Recovery and one version of Planning. Specifically, we gathered
naive speakers' productions of sentences which varied the length of three post-verbal
arguments using a speaker/listener conversation task. The location of intonational
boundaries in the sentences was determined according to ToBI-labeled break indices, and
by acoustic measures of duration of phonological phrase-final words and following
silence. The predictions of the LRB were evaluated in a series of analyses, which will be
described below and referred back to in a discussion of the findings of the experiment.
The results of the experiment support a model of intonational boundary
production based on two factors: Recovery and Planning. Specifically, the probability of
a boundary at a specific location in a sentence increases with the size of the material a
speaker has recently completed and with the size of the material the speaker is going to
produce. Moreover, the accumulated results support a version of Recovery based on the
size of the most recently completed syntactic constituent (Semantic Grouping) over
versions which quantify Recovery in terms of Integration Distance or the distance back to
the last boundary (Incremental). Semantic Grouping was supported by analyses of
variance conducted at critical points across the sentences, and by regression analyses.
Targeted analyses failed to support specific predictions of Integration Distance, and
regression models that used Integration Distance did not account for speaker behavior as
well as Semantic Grouping did. Finally, the Incremental version made some successful
predictions when boundaries were quantified in one way, but failed to correctly predict
speaker behavior when boundaries were quantified differently, and failed to account for
as much variance as Semantic Grouping in a series of regression analyses. These results
support Watson and Gibson's original formulation of the LRB, which we restate in (7),
with Recovery changed to Semantic Grouping.
49
(9) Left Constituent / Right Constituent Boundary (LRB) Hypothesis: Two independent
factors are summed to predict the likelihood of an intonational boundary at a
phonological phrase boundary in production.
1) Semantic Grouping (LHS): The number of phonological phrases in the largest
syntactic constituent that has just been completed. A constituent is completed if it has no
rightward dependents.
2) Planning (RHS): The number of phonological phrases in the largest upcoming
syntactic constituent if it is not an obligatory argument of the most recently processed
constituent.
Summary of analyses
The first set of analyses consisted of a series of analyses of variance where the dependent
measure was the proportion of boundaries which speakers produced at phonological
phrase boundaries in the sentences. Boundaries in this set of analyses were quantified in
two ways:
As places where the ToBI coders indicated boundaries of strength '3' (intermediate
boundary) or '4' (full intonational boundary).
As the sum of the duration of the phonological phrase-final word and any following
silence.
The second set of analyses consisted of a series of analyses of variance where the
dependent measure was the proportion of boundaries which speakers produced at
phonological phrase boundary in the sentences. Boundaries in this set of analyses were
quantified in two ways:
As places where the ToBI coders indicated boundaries of strength '4' (full intonational
boundary).
As the duration of any silence which followed a phonological phrase-final word.
The third set of analyses consisted of a series of regressions, in which we compared the
fit of the data to models which consisted of the planning component paired with one of
the four versions of recovery: (1)semantic grouping, (2) integration distance, (3)
incremental based on the labels of ToBI labeler 1.or (4) incremental based on the labels
of ToBI labeler 2. That is, we compared how often speakers placed boundaries in the
locations that each of the models predict. In addition, we computed the model fit on a
trial-by-trial basis, as well as over the means of the conditions. Boundaries for these
analyses were once again quantified as places where the ToBI coders indicated
boundaries of strength '3' (intermediate boundary) or '4' (full intonational boundary).
The fourth set of analyses consisted of a second set of regression models in which we
again compared the fit of the data to models which consisted of the planning component
paired with one of the four versions of recovery. In this set of analyses, however,
boundaries were quantified as places where the ToBI coders indicated boundaries of
strength '4' (full intonational boundary).
The fifth and final set of analyses was a series of regression equations where we
compared the fit of the data to models which consisted of the planning component paired
with one of two versions of recovery (semantic grouping or integration distance). Here,
however, the regression model was no longer predicting ToBI labels but rather the
duration of phonological phrase-final words and any following silence.
50
Summary offindings
Support for Recovery:
Evidence consistent with all recovery theories:
1. Analysis 1a demonstrated that speakers placed more boundaries after long direct
objects than short direct objects and that speakers placed more boundaries after
long indirect objects than short indirect objects.
2. Analysis lb demonstrated that phrase-final words which completed long direct
objects were longer, and more likely to be followed by silence, than words which
completed short direct objects. In addition, this analysis demonstrated that
phrase-final words which completed long indirect objects were longer and more
likely to be followed by silence than words which completed short indirect
objects.
3. Analysis 2a demonstrated that speaker placed more boundaries after long direct
objects than short direct objects, and that speakers were more likely to place
boundaries after long indirect objects than short indirect objects
4. Analysis 2b demonstrated that longer silences followed long direct objects than
short direct objects.
Evidence that helps decide among recovery theories:
Semantic Grouping
The Semantic Grouping version of Recovery was supported by the main effects of length
of direct object, indirect object, and modifier demonstrated in analyses la, lb, 2a, and 2b.
Semantic Grouping was also supported over the other two versions of Recovery by the
regression models tested in analyses 3, 4, and 5. In each of these analyses, the models
based on Semantic Grouping accounted for more variance in the speaker production data
than the other two models. This result held when the models were run on the data on a
trial-by-trial basis and when the models were run on the means of each condition.
Incremental Distance
Analysis 1a provided support for the Incremental Distance version of Recovery in two
ways: by a two-way interaction between the length of the direct object and the length of
the indirect object, and by a three-way interaction between length of direct object, length
of indirect object, and length of modifier. The two-way interaction suggests that the
placement of a boundary at an early point in the sentence influences the probability of a
boundary later in the sentence. Specifically, if a speaker had not placed a boundary after
a short direct object, s/he would be more like to place a boundary after a short indirect
object, a result that is in contrast to the predictions of both the Semantic Grouping and
Integration Distance versions of Recovery. The three-way interaction suggests that,
when the modifier is long enough, it will be preceded by a boundary, regardless of
whether or not a boundary has recently been produced.
Integration Distance:
In analysis 1a, the Integration Distance version of Recovery failed in two important ways,
as evidenced by a null effect in one critical comparison, when an effect was predicted,
and a strong effect in another comparison where a null effect was predicted. Specifically,
speakers placed more boundaries at the syntactic boundary between the indirect object
and the modifier when integration distance was matched, but the length of the
immediately preceding constituent (i.e. the indirect object) was varied. Moreover, when
integration distance was varied, but the length of immediately preceding constituent (i.e.
the indirect object) was held constant, the incidence of boundaries was the same.
In analysis 2a the Integration Distance version of Recovery was again not supported in
two ways: Speakers' boundary production did not change when the distance back to the
verb was manipulated but the size of the most recently-completed constituent was the
same; and speakers' boundary production increased when the size of the completed
constituent increased but integration distance back to the verb was held constant.
Support for Planning:
1. Results from analysis 1a indicated that speakers were somewhat more likely to
place boundaries before long indirect objects than short indirect objects. In
addition, they were more likely to place boundaries before long modifiers than
short modifiers.
2. Results from analysis lb demonstrated that speakers produced significantly longer
phrase-final words and measurable silence preceding long modifiers than short
modifiers.
3. Results from analysis 2a indicated that speakers were more likely to place
boundaries before long indirect objects than short indirect objects and that
speakers were more likely to place boundaries before long modifiers than short
modifiers.
4. Results from analysis 2b indicated that speakers placed longer silences before
long modifiers than short modifiers.
5. Results from analyses 3, 4, & 5 indicated that in every model tested, the Planning
component of the model accounted for a significant amount of variance in speaker
boundary placement.
In addition to offering insight into the processes that underlie spoken language
production, the results of the current study motivate us to reiterate a point made by
Watson & Gibson which may be able to explain some of the discrepancies between
previous investigations of the relationship between syntactic structure and speech
planning (cf. Watson & Gibson, 2004 for similar arguments). For example, although
Kraljic and Brennan (2005) found evidence that speakers will prosodically disambiguate
a syntactically-ambiguous sentence whether or not they are aware of the ambiguity,
Snedeker and Trueswell (2003) found that speakers only prosodically disambiguated
sentences if they were aware of the ambiguity. We suggest that the important difference
between the two studies, which Kraljic and Brennan point out, is the length of the stimuli
used in each. Whereas the sentences that Snedeker and Trueswell used were comprised
of three phonological phrases (e.g. "Tap the frog with the feather"), the sentences in
Kraljic and Brennan's studies were comprised of four phonological phrases (e.g. "Put the
dog in the basket on the star").
Both the modifier attachment production and the goal attachment production of
"Put the dog in the basket on the star" result in places in the sentence where the LRB
weight is high enough to induce speakers to place boundaries as a result of constraints on
speech production. When the modifier attachment of "on the star" is intended, the
semantic groupingcomponent of the LRB has a weight of one ("the dog"), while the
planningcomponent of the LRB has a weight of two ("in the basket on the star"),
resulting in a total LRB weight of three. Similarly, when the goal attachment of "on the
star" is intended, the semantic groupingcomponent of the LRB has a weight of two ("the
dog in the basket"), while the planningcomponent of the LRB has a weight of one ("on
the star"), resulting in a total LRB weight of three. In contrast, both the modifier and goal
attachment productions of "Tap the frog with the feather," result in LRB weights of only
two. When the modifier attachment of "with the feather" is intended, a boundary would
be expected between "Tap" and "the frog," but the LRB weight of two may not be large
enough to induce the speaker to place a boundary in this location. Similarly, when the
instrument attachment of "with the feather" is intended, a boundary would be expected to
produced between "the frog," and "with the feather." Once again, the total LRB weight
at this location may not be enough to induce the speaker to insert a boundary.
However, the additional pressure of having a listener understand an ambiguous
sentence was enough to induce the speakers to place disambiguating boundaries in the
sentence. That is, the speakers' knowledge of the ambiguity, and their need to
communicate the ambiguity to their listener, can induce them to place a boundary at a
point in the sentence even if the LRB weight is lower than three phonological phrases.
The assertion that there is a threshold LRB weight at which a speaker must place a
boundary, and that the desire to communicate to a listener can lower this threshold and
induce boundaries, can be tested in future work using sentences in which the same types
of constituents differ systematically in length.
A final important contribution of this study is the demonstrated relationship between the
quantitative and qualitative evaluations of the speakers' data. That is, we observed high
correlations between the acoustic measures of word duration and/or silence and the
perceptual ToBI labels, indicating, as previously noted, that phrase-final lengthening and
silence are strong cues to the presence of an intonational boundary. These correlations
are especially noteworthy considering that the words which are being compared have
very different segmental characteristics and are produced by different speakers, due to the
fully between-subject design of the experiment. Although the acoustic results did not
always mirror the ToBI codings, it is plausible that we would not observe all the effects
in the acoustic data that we observed for the perceptual data, because the duration and
silence are only two of the many acoustic cues that gives rise to the perception of a
boundary. The relationship that we did observe, however, suggests that with more
research into the acoustic cues that give rise to the perception of intonational boundaries,
future investigations of phrasing and syntactic structure may be undertaken without
necessitating the involvement of multiple, highly-trained prosodic annotators.
Chapter 4
Chapter 2 demonstrated that coders achieved very high agreement on the application of
binary prosodic categories to a large corpus of speech. These results suggest that the
binary categories of both prosodic annotation systems (ToBI and RaP) are good
reflections of the prosodic features that speakers and listeners are using to communicate.
However, labeler agreement decreases substantially for categories with more than two
levels, suggesting that these multi-level categories may not be readily applied to the
normal speech system, and that prosodic annotation systems are not accurately capturing
important categories of prosody.
The goal of the set of studies described in this chapter was to undertake a
methodologically sound exploration of the relationship between the acoustics of words
and their information structure. Some of the specific questions we will address are:
1. What are the acoustic features that indicate an entity as accented?
2. Do the acoustic features of entities which are widely focused differ from entities
which are deaccented?
3. Are accents acoustically differentiated by speakers on the basis of the type of
information status of the accented entity, specifically, for entities which are new
or contrastive in the discourse?
4. Are certain types of information status differentiated more readily than others?
5. If (4) is true, under what circumstances will speakers differentiate information
status?
6. Is there a reliable relationship between the extent to which a speaker acoustically
encodes given/new or new/contrastive distinction and a listener's ability to
resolve the information structure?
The first section of this chapter will describe a series of studies which
demonstrate that listeners are sensitive to information structure as indicated by prosodic
features. Next, we will describe early attempts at characterizing the acoustic
characteristics of different information status categories.
Interpretation ofpitch accents
This section will recount studies which have shown how naYve listeners are able to
recognize and utilize accents to decide which information is new or given in a sentence,
and that this information is available to listeners immediately.
Most and Saltz (1979) used a question-choice paradigm to investigate whether
listeners interpret stress as signaling new information. Listeners heard sentences like The
PITCHER threw the ball or The pitcher threw the BALL, where words in caps indicate
words associated with pitch accents, and were asked to write a question for which the
sentence would be an appropriate reply. In 68% of responses, the listeners' questions
semantically focused the stressed element (e.g. Who threw the ball? in response to The
PITCHER threw the ball), demonstrating that listeners are interpreting accent placement
as indicating the new information in an answer. Although these results are suggestive, no
quantitative definition of accent in given in this paper, and consequently, there is no way
to know what acoustic or prosodic cues the listeners were interpreting as signaling accent.
Bock and Mazzella (1983) also asked whether accents signal new information, but
also investigated the consequences of accent placement for on-line comprehension.
Listeners heard a context sentence followed by a target sentence, and were instructed to
press a button when they understood what the target sentence meant. The four conditions
are presented in Table I. The critical manipulations were whether a) the context sentence
semantically focused an element by the placement of an accent, and b) whether the accent
in the target sentence was on the item which was semantically focused by the context
sentence.
Condition
1. Appropriate accent
2. Inapproriate accent
3. No context accent
4. Control
Context Sentences
ARNOLD didn't fix the radio.
Arnold didn't FIX the radio.
Arnold didn't fix the radio.
Arnold didn't fix the radio.
Target Sentences
DORIS fixed the radio.
DORIS fixed the radio.
DORIS fixed the radio.
Doris fixed the radio.
Table 1: Sentence Setfrom Bock andMazzella (1983) Experiment 1
The results demonstrated that listeners understoond the target sentence most quickly
when it had been preceded by a context sentence with an appropriately placed accent
(condition 1). The fact that condition 1 resulted in faster comprehension times than the
"No context accent" condition (4) demonstrates that the semantic focusing of the element
that will be contrasted in the target sentence facilitates listeners' comprehension of the
target sentence. Once again, there is no discussion of the quantification of accent in this
experiment, save the agreement of the accent producer and one other listener. Therefore,
this paper leaves open the question of how accents are in fact realized such that they can
facilitate listener comprehension of contrastive information.
Most and Saltz (1979) and Bock and Mazzella (1983) demonstrated that the
presence of accents can facilitate listeners' comprehension. There is also evidence from
Terken and Noteboom (1987) that listeners' comprehension can be facilitated by the
absence of accents. In their experiment, listeners heard spoken descriptions of changes
that were taking place in a visually presented configuration of letters. The listeners' task
was to verify the spoken descriptions, like "the p is on the right of the k," as quickly as
possible. Previous descriptions were manipulated so that both the 'p' and the 'k' were
either given or new in the target description. In accordance with prior results, Terken and
Noteboom found that listeners were faster to verify new information that was accented
than new information that -wasdeaccented. In addition, they demonstrated that given
information that was deaccented was more quickly verified than given information that
was accented, indicating that listeners are able to use deaccentuation as a cue to the
discourse status of the deaccented element.
Birch and Clifton (1995) reported results similar to those of Terken and
Noteboom. In a set of four experiments, Birch and Clifton operationalized accents
according to the ToBI system of prosodic annotation. They had subjects make speeded
judgments of the appropriateness of answers to questions. They found that listeners were
faster to indicate that a sentence was an appropriate response to a question when what
was new in the sentence was accented or what was given in the sentence was deaccented.
Although the studies described above by Bock & Mazzella (1983), Terken and
Noteboom (1983), and Birch and Clifton (1995) demonstrate that the appropriateness of
accents placement can have an early effect on listeners' comprehension, they leave open
55
the question of when this referential information is available to listeners. Dahan,
Tanenhaus, & Chambers (2002) conducted an eye-tracking study which showed that
listeners can use accent information immediately to decide the identity of a temporarily
ambiguous word by determining whether it was given or new in the discourse. The
researchers presented subjects with a visual display containing four objects and four
geometric shapes. The critical objects in the test trials were two nouns that shared the
same first syllable (e.g. candle, candy) or onset and nucleus (e.g. bell, bed). Listeners
followed spoken instructions like the following to move these items around the display:
1
2
a. Put the candle below the triangle.
b. Put the candy below the triangle.
a. Now put the CANDLE above the square
b. Now put the candle ABOVE THE SQUARE.
The eye-tracking data demonstrated that when subjects heard sentence la
followed by 2a, they were more likely to look at the competitor (candy) than the target
(candle) than when sentence la was followed by 2b. That is, subjects interpreted the
accent on "candle" in sentence 2a as indicating that the word with which it was
associated was new to the discourse. Conversely, when subjects first heard sentence 1b,
they were more likely to look to the competitor (candle) when it was unaccented (2b)
than when it was accented (2a). In this case, subjects interpreted the deaccenting of
"candle" as indicating that it was given in the discourse.
The studies described above all demonstrate that both the presence and absence of
accents on semantically focused sentence elements affect listeners' comprehension of the
discourse. One question left open by the studies described above, however, is whether
different types of accents have different meanings. Most and Saltz (1979) and Bock and
Mazzella (1983), and Terken and Noteboom (1983), provide no description of the
phonological or acoustic realization of the accents in their studies. Birch and Clifton
(1995) indicate that all accents in their four experiments were realized with a L+H*
accent, according to the ToBI conventions, which will be described below; however, in
the ToBI framework, the L+H* accent is hypothesized to indicate contrastive information
(Pierrehumbert & Hirschberg, 1990), and in Birch and Clifton's experiments, all accents
indicated new information. Dahan, et al. (2002) also report the phonological categories
of their accents, but did not control for the type of pitch accent that was associated with
the target noun in their study. In fact, the ToBI codings they report indicate that out of 24
cases, the accent was H* in 15 cases, and L+H* in the remaining 9. The following
section will discuss hypothesized semantic categories of accents, and experimental work
designed to investigate the reality of these categories, both acoustically and perceptually.
Categories ofpitch accents
There has been extensive debate throughout the history of research on intonation
about the extent to which different accents signal different types of information. There
are two major camps in this debate: those who argue that different pitch accents indicate
different semantic categories (Pierrehumbert & Hirschberg, 1990; Gussenhoven, 1983),
and those who believe that a single accent can indicate one (or more) meanings
depending on the context in which it occurs (Bolinger, 1961; Cutler, 1977).
The most attention has been paid to the question of whether different categories of
accent signal new and contrastive information. Bolinger (1961) argues that the meaning
of a particular accent is defined by the context in which it occurs; that there is no unique
contrastive intonation, as defined by pitch. Cutler (1977), likewise, claims that "the
attempt to extract from [intonation contours] an element of commonality valid in all
contexts must be reckoned a futile endeavor" (p. 106).
In contrast to the position advocated by Bolinger and Cutler, however, is that
taken by Gussenhoven (1983) and Pierrehumbert & Hirschberg (1990). Gussenhoven,
having defined accents asfalls, rises, andfall-rises,argues that afall means that the
speaker is adding the accented item to the discourse, whereas thefall-rise indicates that
the speaker is selecting the accented item from those already in the discourse.
Gussenhoven's position is similar, then, to Pierrehumbert and Hirschberg (1990), who
propose a compositional semantics of intonational meaning where pitch accents, phrase
accents, and boundary tones combine to form the meaning of a sentence. With regard to
pitch accents, they claim that a high tone on a stressed syllable (H*) signals new
information and a high accent preceded by a low (L+H*) signals contrastive or corrective
information They suggest that a speakers' use of H* indicates that the proposition
including the H* should be added the listener's beliefs. In contrast, a speaker's use of
L+H* indicates that the s/he intends that "the accented item-and not some alternative
related item-should be believed" (p.296).
What is missing from the above debate about whether or not different accents can
convey meaning differences irrespective of the context in which they occur is empirical
investigation into the acoustic realization of accents associated with new information and
those indicating contrast.
Acoustic realization of accents
This section will discuss work aimed at defining accents acoustically, and,
furthermore, whether there are acoustic differences between accents that have been
proposed to have different meanings.
Previous work has suggested that a large variety of acoustic measures indicate the
presence of accents in English. The list of proposed factors includes: pitch (Lieberman,
1960; Cooper, Eady & Mueller, 1985; Eady and Cooper, 1986), duration (Fry, 1954), and
energy (Bolinger, 1958; Kochanski, Grabe, Coleman, & Rosner, 2005).
Early studies on the acoustic correlates of stress focused on lexical stress. Both
Fry (1954) and Lieberman (1960) investigated this question by measuring acoustic
features of words whose part of speech varies with stress. For example, con-TRAST is a
verb, while CON-trast is a-noun. In both production and perception studies, Fry found
that intensity and duration of the vowel of the stressed syllable contributed most strongly
to the noun-verb disambiguation, such that stressed vowels were produced with a longer
duration and a greater intensity than non-stressed vowels. Lieberman replicated Fry's
results by demonstrating that stressed syllables had higher amplitudes and longer
durations than their non-stressed counterparts. Moreover, he demonstrated that the
fundamental frequency of stressed syllables was higher than non-stressed syllables.
Finally, Lieberman proposes a decision algorithm that can discern, with 99.2% accuracy,
which syllable of a two-syllable word is stressed by combining information about FO,
amplitude, and duration.
Studies investigating acoustic features of phrasal stress have investigated similar
variables. Cooper, et al. (1985), for example investigated the role of duration and
fundamental frequency in speakers' production of focus. In this study, they manipulated
the position of focus in a target sentence like: Chuck liked the present that Shirley sent to
57
her sister. They did this by manipulating which element in the sentence contrasted with
prior discourse, as the example in (2) demonstrates:
A. Did William or Chuck like the present that Shirley sent to her sister?
(2)
B. Did Chuck like the letter or the present that Shirley sent to her sister?
C. Did Chuck like the present that Melanie sent to her sister or the one that
Shirley sent?
D. Did Chuck like the present that Shirley sent to her sister or the one she sent to
her brother?
With regard to duration, Cooper, et al. demonstrated that the duration of the
focused word was significantly greater than when it was not focused. The last word of
the sentence (e.g. sister) showed a significantly smaller increase in duration due to focus,
but this may be because all sentence-final words are lengthened, thus washing out
duration differences due solely to focus. There was no effect of deaccenting on duration,
in that words which occurred after the sentence focus were not shortened relative to nonfocused words in the same position.
The effects of fundamental frequency (FO) were similar to those of duration in
that all but the sentence-initial focused word (e.g. Chuck) exhibited a higher FO than nonfocused words in the same position. In contrast to the duration results, however, there
was evidence of deaccentuation in the FO results, such that words which occurred after
the sentence focus were produced with a lower mean FO than non-focused words in the
same sentence position.
In a second study, Eady & Cooper (1986) again investigated the role of FO and
duration in the production of stress in sentences like: The ship is departingfromFrance
on Sunday. Contrary to the previous experiment, however, they defined the focused
element of a sentence in terms of what was focused by a wh-question, as in (3). In this
way, the words which were stressed in this experiment would be stressed by virtue of
being new information, and not be virtue of being contrastive, as they had been in the
previous study.
(3)
A. What is happening?
B. What is departing from France on Sunday?
C. On what day is the ship departing from France?
The results from this experiment were quite similar to the previous experiment in
which speakers produced stress on contrastive elements. First, stressed elements were
produced with longer durations in both initial and final positions. Second, focused words
were realized with higher FO than non-focused words. Again, there was evidence of
deaccentuation such that post-focal words were realized with a lower FO than neutral
words.
From the results of these two experiments, Eady and Cooper (1986) conclude that
"the type of sentence focus used [in the second study] has essentially the same effect on
acoustical attributes as does contrastive stress" (p.408), suggesting that there is no
acoustic difference between accents realizing new and contrastive stress. However, this
claim is hard to evaluate for two reasons: First, these results come from two
independently conducted experiments, and it is not possible to compare speaker behavior
across experiments; second, the authors limited their analyses to only select acoustic
measurements, and differences would have been observed if they had looked at additional
58
parameters. The next section will report results of experiments in which new and
contrastive stress are compared within the same experiments.
One caveat of both studies conducted by Eady and Cooper is that, although they
were investigating the productions of multiple speakers, they did not include all speakers
in their analyses. They selected speakers for analysis based on the extent to which the
speakers produced accents in the appropriate places, as determined by one of the authors
and another listener.
Although all of the studies described in this section thus far have found evidence
of FO differences between stressed and unstressed words, a recent study has called into
question the importance of FO in determining the perception of accents. Kochanski, et al.
(2005) trained a classifier to recognize pitch accents which had been hand-labeled in a
corpus of read and spontaneous speech, using five acoustic measures: loudness, duration,
aperiodicity (a measure of the periodicity of the waveform ranging from 0 to 1), spectral
slope (an acoustic correlate of the 'breathiness' of the speech), and fO. They found that
loudness was the best predictor of accents, regardless of speaking style or dialect.
Moreover, FO was the worst of the five predictors in determining whether a syllable was
accented.
The studies described in this section provide evidence that FO, intensity
(loudness), and duration all contribute to some degree to the perception of accent. The
following section will explore a series of experimental studies intended to explore
whether these acoustic features are different depending on whether the accent they define
is signaling "new" or "contrastive" information.
New vs. contrastive accents
The first part of this section will review recent work exploring whether new and
contrastive stress are perceived categorically. The second section will review work
exploring whether these accents are produced categorically.
Perception
Bartels and Kingston (1994) synthesized a continuum of stimuli intended to vary
between H* and L+H* by independently varying four acoustic characteristics of the
target accent. They then presented sentences containing these synthesized accents to
naive listeners and asked them to make a meaning judgment, in order to test the
hypothesis that listeners would interpret entities accompanied by an H* as new to the
discourse but interpret entities accompanied by L+H* as contrastive with information
already in the discourse (Pierrehumbert & Hirschberg, 1990). From their results, they
conclude that the most salient cue to contrastiveness is the height of the peak of a high
accent, such that a higher peak indicates greater contrast. Two secondary cues to
contrastivensss were (a) the depth of the dip preceding the peak, such that deeper dips
indicated greater contrastiveness and (b)the timing of the peak, such that early peaks
were interpreted as more contrastive than late peaks. Importantly, there was no clear
evidence of a categorical boundary between L+H* and H* in terms of meaning
differences. That is, it was not the case that L+H* consistently signaled contrastive
information while H* consistently signaled new information. Bartels and Kingston
suggest that the main finding that a higher peak indicates greater contrastiveness supports
the possibility that contrastive accents are merely more salient versions of new accents.
Watson, Tanenhaus, & Gunlogson (2004) used an eye-tracking paradigm similar
to that utilized by Dahan, et al. (2002) to explore whether listeners immediately interpret
59
H* and L+H* categorically, with H* indicating new material and L+H* indicating
contrastive material. Listeners heard the directions in (1), while interacting with the
items in a computerized display. The target item (e.g. camel/candle) in sentence Ic was
crossed with the type of accent aligned with it in a 2x2 design.
(1)
a. Click on the camel and the dog.
b. Move the dog to the right of the square.
c. Now, move the camel/candle below the triangle.
L+H*/H*
Eye-tracking results demonstrated that listeners were more likely to look to the
contrastive referent (the camel) when they heard the L+H* accent than to the new
referent (the candle), suggesting that listeners quickly interpreted the L+H* accent as
contrastive. In contrast, listeners did not look more quickly at the new referent (the
candle) when it was indicated with a H* accent, indicating that H* is not preferentially
treated as signaling new information. Although these results suggest that L+H* is
preferentially interpreted as interpreting contrastive information, they do not indicate that
H* is preferentially treated as signaling new information. They support the notion,
rather, that L+H* has preferentially meaning, but that H* could indicate either meaning,
and, therefore, that there does not exist a categorical distinction between the two accents.
Taken together, these perception studies suggest that listeners can perceive
differences between accents depending on the intended meaning. However, the
difference between accents that signal new and contrastive meanings does not appear to
be categorical, such that one accent preferentially signals new information and another
signals contrastive information. These data also suggest that differences between new
and contrastive accents are determined by the height of the pitch peak on the accented
syllable, rather than a difference in the overall shape of the accent.
Production
Ito, Speer, and Beckman (2004) developed a novel paradigm for eliciting
spontaneous productions which would vary in their information structure, including not
only elements that were new or given with respect to previous context, but also items
which were contrastive with previous discourse. They found that given adjectives often
received a pitch accent, whereas given nouns were less likely to be accented.
Conversely, both adjectives and nouns which were used contrastively (e.g. "green candy"
preceded by "beige candy" or "green candy" preceded by "green house") were more
likely to be produced with a L+H* accent than when used non-contrastively. These data
give some suggestion that speakers use different accents to indicate items which are new
vs. items which are contrastive. One advantage of this study over other production
studies is the inclusion of data from sixteen speakers, ensuring that observed differences
are not merely due to idiosyncrasies of individual speakers; however, the study is limited
in two important ways: First, there is no explanation of the acoustics of the accents; only
the ToBI labels are reported; second, no statistical measures of the reported differences
are provided.
Krahmer & Swerts (2001) also investigated the shape of contrastive accents and
how they differ (if at all) from new accents. To do this, they had eight subjects engage in
a game task, where they produced noun phrases (adjective + noun) in one of four
60
discourse contexts: (1) both new (NN), (2) both contrastive (CC), (3) contrastive
adjective / given noun (CG), (4) given adjective / contrastive noun (GC). Two raters
identified the location of accents and found that in cases where both adjective and noun
were new or contrastive, speakers tended to place accents on both adjective and noun.
When only one of the two was contrastive, the contrastive element tended to receive a
pitch accent while the non-contrastive element was deaccented.
To investigate whether contrastive accents differ from new accents, Krahmer and
Swerts presented the utterances of two speakers from the production study to listeners.
They played the productions in one of two contexts: either the adjective and noun
together, or each in isolation. The listeners' task was to say which item in a pair was
more prominent. The results indicated that listeners perceived contrastive accents as more
prominent than new accents when the accents were presented in context; when presented
in isolation, there is no clear pattern of difference between contrastive and new accents.
Krahmer and Swerts interpret these results as indicating that, although there is no
categorical difference between new and contrastive accents, the difference between new
and contrastive accents can readily be determined from the context in which these accents
are produced. Unfortunately, Krahmer & Swerts present results from only two of eight
speakers, and include no explanation for the selection of those two speakers.
Calhoun (2005), like Krahmer and Swerts, also addressed the possibility the
differences between new and contrastive accents can be realized non-locally; that is, that
the acoustic features of words in the local domain of an accented word can vary
systematically when the type of accent on the focused word varies even when the
acoustics of the focused word do not differ. She explored this hypothesis in a pilot study
where she compared version of the phrase "when that moves the square" where the
information status of square varied between given, new and contrastive with respect to
previous discourse. She first demonstrated that local acoustic cues differ significantly
depending on discourse status in that a regression model demonstrated that the mean FO
of the, the mean FO of square, and the duration of square all significantly predicted the
information status of square. Second, she found that the non-local context of square
could also predict it's discourse status, such that the FO and intensity of the phrase that
moves were also significant predictors of the topic status of square. Finally, she
demonstrated that the FO and intensity of square relative to the FO and intensity of that
and moves could also predict the discourse status of square. Taken together, these results
suggest that the topic status of a word can be reflected in the acoustics of the word's
entire phrase, and not just by acoustic measures of the word itself.
Finally, Calhoun (2003) also attempted to discern a reliable difference between
H* and L+H*. In a production study, she had one speaker produce sentences in which
the two accents had been disambiguated by context. She first identified where the
speaker had placed accents, and, for each accent, measured a series of acoustic features.
The results of this portion of the experiment demonstrated that the two accents were most
strongly differentiated by (a) the alignment of the pitch minimum preceding the high and
(b) the height of the high maximum. A subsequent perception study, in which listeners
chose their preferred rendition of an accent in a semantic context, demonstrated that only
the height of the pitch maximum mattered to the disambiguation of new and contrastive
accents.
As in the perception studies, the production studies reported here also suggest that
speakers can differentiate accents on the basis of meaning. They provide additional
support for the hypothesis contrastive accents are indicated by a higher pitch peak, and
that contrastive accents are realized with greater intensity than new accents.
Current Study aims
We designed the current set of studies to address limitations of the studies
described above and to investigate the function of acoustic features in the determination
of information status.
Limitations of previous studies
We first identified several methodological limitations of previous investigations
of the relationship between acoustic measures and information status. With regard to the
perception of prosody, most studies use trained speakers to produce their materials
(Calhoun, 2003; Birch & Clifton, 1995, Most & Saltz, 1979; Bock & Mazzella, 1983).
These speakers are presumably fully aware of the experimental aims, and therefore aim to
produce maximally different prosody for different semantic conditions. This approach is
problematic, because the obtained results from these studies may not generalize to all
speakers or all perceivers. Even when production studies do employ naive speakers, they
do not use all of the speakers. Several previous experiments have excluded speakers'
data from analysis for not producing accents consistently (Eady & Cooper, 1986), or with
no explanation (Krahmer & Swerts, 2001).
The current studies attempt to account for these stated limitations in several ways:
First, the speakers in the current studies were not trained speakers, so any differences in
their productions should be naturally occurring. Second, we did not exclude speakers
from our experiments on the basis of whether or not they behaved as we hypothesized
they should (i.e. placed accents in particular places). We excluded only subjects who
were not native English speakers, who did not take the task seriously, or who were not
recorded well.
A limitation of previous studies of the perception of prosody is that, rather than
report acoustic measures, they often report only the ToBI annotations of their materials
(Birch & Clifton, 1995; Ito, et al., 2004). Investigations of inter-annotator agreement in
ToBI suggests that H* and L+H* are the most often confused in the ToBI system (Syrdal
& McGory, 2000) and are often collapsed in these studies (Pitrelli, et al., 1994; Yoon, et
al., 2004). Therefore, it is difficult to interpret results of studies which are based on the
difference between H* and-L+H* without any reporting of the acoustic differences
between these categories. To avoid the confusion inherent in ToBI labels, we report
acoustic measures which differ across conditions, and not just ToBI labels, in order to
avoid confusion about what these labels might mean.
Finally, the current studies attempt to improve on the task demands of previous
studies. Previous studies asked listeners to make judgments about which of two stimuli
was more prominent (Krahmer & Swerts, 2001), or what accent is acceptable in a
particular context (Birch & Clifton, 1995). Our producers and perceivers were engaged
in a meaning task; the perceivers were trying to communicate a particular meaning of a
sentence, and our dependent measure was the perceiver's semantic interpretation of the
sentence. This approach is an improvement over previous methods because it relates
differences in accent types to differences in meaning, rather than simply to perceptual
differences.
The details of the current method for Experiments 1-3 were as follows: Each
speaker in the experiment produced the same sentence under seven different information
status conditions, which were manipulated by the setup question that the speaker was
answering with each question. These conditions were produced by putting either new or
contrastive focus on either the subject, verb, or object. From the speaker's production, a
Listener chose from seven possible sentences which one s/he though the speaker was
answering. We performed acoustic analyses on the productions that speakers were able
to correctly categorize, assuming that those productions bore the correct cues to
information status and focus location.
Experiment 1
Method
Participants
We recorded 16 pairs of subjects for this study. Six speakers were excluded from
analysis for the following reasons: One speaker was not a native American English
speaker, two speakers were too quiet to be acoustically analyzed, and three did not take
the task seriously, often laughing during their productions. Each subject received ten
dollars/hour for his/her participation in both experimental roles (speaker and listener).
Materials
Each trial consisted of a set-up question and a target sentence. The target
sentence could plausibly answer any one of the seven set-up questions, which served to
focus different constituents on the sentence. Two factors were manipulated: (1) the
constituent in the target sentence that was focused by the question (subject, verb, object);
and (2) the discourse status of the focused constituent (new, contrastive). In addition, we
added an additional condition which focused the entire sentence (i.e. What happened?),
for a total of seven conditions. We included this condition in order to see whether
accents in any the narrow-focus conditions differed from those in the wide-focus
condition.
The words in the target sentences were chosen so that they could be compared
across items and to aid in the extraction of acoustic features. To this end, all subject
names and object noun phrases (NPs) were two-syllable names/words with first-syllable
stress, comprised of sonorant phonemes, such as "Damon" and "omelet." All verbs were
one-syllable, comprised mQstly of sonorant phonemes, such as "fried."
We constructed 14 sets of all 7 conditions, resulting in 98 experimental items.
Each subject pair was presented with every item, resulting in a full within-subjects
design. A complete item is presented in Table 1. All materials can be found in Appendix
A.
Condition
Status
Focused
Argument
Setup Question
1
New
wide
2
New
S
3
New
V
Who fried an omelet yesterday?
What did Damon do to an omelet
yesterday?
4
5
New
Contrastive
0
S
What did Damon fry yesterday?
Did Harry fry an omelet yesterday?
What happened yesterday?
Target
Damon fried an
omelet yesterday.
Damon fried an
omelet yesterday.
Damon fried an
omelet yesterday.
Damon fried an
omelet yesterday.
No, Damon fried an
6
Contrastive
V
Did Damon bake an omelet yesterday?
7
Contrastive
0
Did Damon fry a chicken yesterday?
omelet yesterday.
No, Damon fried an
omelet yesterday.
No, Damon fried an
omelet yesterday.
Table 1: Example item from Experiment 1
Procedure
The experiment was conducted using Linger (2.92), a software platform for
language processing experiments. Linger was written and designed by Doug Rohde, and
can be downloaded at http://tedlab.mit.edu/-dr/Linger/. Two participants were included
in each trial, and sat at computers in the same room such that neither could see the others'
screen. One participant was the speaker, and the other was the listener. The speakers
were instructed that they would be producing answers to questions out loud for their
partners (the listeners), and that the listeners would be required to choose which question
the speaker was answering from a set of seven choices.
Each trial began with the speaker being presented with the question on the
computer screen to read silently until s/he understood it. The speaker then saw the
answer to the question, accompanied by a reminder that s/he would only be producing the
answer aloud. Following this, the speaker had one more chance to read the question and
answer, and then was instructed to press a key to begin recording, and another key to stop
recording.
The listener sat at another computer, and pressed a key to see the seven questions
that s/he would have to choose his/her answer from. When s/he felt familiar with the
questions, s/he told the Speaker s/he was ready. After the speaker produced a sentence
out loud for the listener, the listener chose the question s/he thought the speaker was
answering.
In early pilots in which there was no feedback for incorrect responses, we
observed that Listeners were at chance to choose the correct question, and that Speakers
were not perceptibly disambiguating the answers for the Listener. In order to remedy this
situation, we introduced feedback for both the Speaker and the Listener such that when
the Listener answered incorrectly, his/her computer emitted an audible buzz. Listeners
who received this feedback were well above chance levels in choosing the correct
answer, presumably because the Speakers were explicitly aware when they had not
provided accurate prosodic-cues to the correct answer.
Data Analysis
We report results based on productions of all speakers in the experiment except
those trials that were not recorded or cutoff for technical reasons, trials that were too
quiet to contribute useful acoustic data, and trials in which the speaker was disfluent, or
used different words from those presented on the screen. Overall, these exclusions
results in 105 of the 980 total trials (9%).
Acoustic Factors
Based on previous investigations of the acoustic correlates of prosodic features,
we chose a series of acoustic measures which we believed would reflect accentuation.
For each word, we obtained measures of the following, using the Praat program
(Boersma & Weenink, 2006):
1. duration: duration, in ms, of each word, excluding any silence before or after the
word.
2. silence: duration, in ms, of any measurable silence following the word, which was
not due to stop closure.
3. duration + silence: the sum of the duration of the word and any following
silence, in ms.
4. mean pitch: The mean FO, in Hz, of the entire word
5. maximum pitch: the maximum FO value (in Hz) across the entire word
6. pitch peak location: a measure between 0 and 1 indicating the proportion of the
way through the word where the maximum FO occurs.
7. minimum pitch: the minimum FO (in Hz) across the entire word
8. pitch valley location: a measure between 0 and 1 indicating the proportion of the
way through the word where the minimum FO occurs.
9. initial pitch: the mean FO value of the initial 5% of the word
10. early pitch: the mean FO value (in Hz) of 5% of the word centered at the point
25% of the way through the word
11. center pitch: the mean FO value (in Hz) of 5% of the word centered on the
midpoint of the word
12. late pitch: the mean FO of 5% of the word centered on a point 75% of the way
through the word
13. final pitch: the mean FO of the last 5% of the word
14. mean intensity: mean intensity (in dB) of the word
15. maximum intensity: the highest dB level in the word
16. minimum intensity: the lowest dB level in the word
17. intensity peak location: a measure between 0 and 1 indicating the proportion of
the way through the word where the maximum intensity (in dB) occurs
18. intensity valley location: a measure between 0 and 1 indicating the proportion of
the way through the word where the minimum intensity (in dB) occurs
19. maximum amplitude: the maximum amplitude (sound pressure in Pascal) across
the word
20. energy: the square of the amplitude multiplied by the duration of the word
21. 1st quarter pitch: The difference, in Hz, between initial pitch and early pitch.
22.2
2d quarter pitch: The difference, in Hz, between early pitch and center pitch.
23.
rd
quarter pitch: The difference, in Hz, between center pitch and late pitch.
24. 4rd quarter pitch: The difference, in Hz, between late pitch and final pitch.
65
Results - Production
Important acoustic factors - ANOVAs
In order to determine which acoustic factors differed across conditions, we
performed a series of 4 x 3 analyses of variance on each acoustic measure. The first
factor, information status, had four levels, corresponding to (1) given, (2) wide focus, (3)
narrow new focus, and (4) narrow contrastive focus. The second factor, sentence
position, had three levels: (1) Subject, (2) verb, and (3) object.
(1) Given words were any words which were not focused in one of the sentences
where either the subject, verb, or object were focused. For example, when the
setup question focused the subject, as in Whofried an omelet yesterday?, or
DidHarryfly an omelet yesterday?, the given words in the answer werefried
and omelet.
(2) Words with wide focus were either the subject, verb, and object when the
setup question focused the entire sentence, as in What happenedyesterday?.
(3) Words with narrow new focus were either the subject, verb, or object when
the setup question imposed new focus on that constituent. For example, if the
setup question was What did Damonfly yesterday?, omelet received new
focus in the answer.
(4) Words with narrow contrastive focus were either the subject, verb, or object
when the setup question imposed contrastive focus on that constituent. For
example, if the setup question was Did Damonfry a chicken yesterday?,
omelet received contrastive focus in the answer.
We performed each ANOVA using both subjects and item as random factors. The results
from the entire series of ANOVAs are presented in Table 1, which indicates for which of
the four information status levels each acoustic measure varied significantly.
Acoustic measure
duration
silence
duration + silence
mean pitch
maximum pitch
pitch peak location
pitch valley location
early pitch
center pitch
maximum intensity
maximum amplitude
energy
1st quarter pitch
2nd quarter pitch
3rd quarter pitch
4th quarter pitch
Information
Status
differentiated
1, 2, 34
1,24, 3
1, 2, 34
1, 2,3, 4
1, 2, 34
1, 234
12, 34
1, 2, 34
1, 2, 3, 4
1, 2, 34
1,2, 34
1, 2, 34
1, 234
123, 4
12, 3, 4
12, 34
66
Table 2: Information status differentiated by individual acoustic features across the
subject, verb, and object.
Acoustic features used in disambiguation
Although the results of the analyses of variance provide some evidence as to
which acoustic features speakers use to disambiguate both the presence or absence of
accents, and the type of information status intended by a particular type of accent, they
cannot answer the question of which acoustic factors were ultimately responsible for
Listeners' condition choices, for two reasons. First, the ANOVAs only indicate acoustic
features which differ between conditions in single sentence locations; they do not provide
any indication of differences across the sentences. Listeners hear each word in the
context of a complete sentence, and presumably interpret the acoustics of each word in
the sentence with reference to the other words. As such, the absolute difference between
words in the same position across different conditions may not matter as much to the
Listener as the relationship between a word of interest (focused or non-focused) and the
other words in the sentence. Second, the ANOVAs do not provide any indication about
the relative importance for the different acoustic features in determining differences in
Listener behavior. In other words, the finding of significant differences between one
acoustic factor across conditions does not necessarily indicate that Speakers and Listeners
are using that acoustic feature to disambiguate between conditions.
Therefore, in order to determine which of the candidate acoustic features actually
mediated the difference between different accent locations or different levels of
information status, we conducted a series of stepwise discriminant analyses on the data
from Experiment 1, where we entered all acoustic features for all words as predictors.
Across all analyses, the acoustic features which consistently resulted in the best
classification of conditions were (1) duration + silence, (2) mean pitch, (3) maximum
pitch, and (4) maximum intensity. From this result, we used only these four factors in the
discriminant analyses that we performed on all subsequent data sets in this chapter.
Following the identification of the critical acoustic measures through the
stepwise discriminant analyses, three subsequent discriminant analyses were conducted to
determine whether the measures of (1) duration + silence, (2) maximum pitch, (3) mean
pitch, and (4) maximum intensity on the three critical words in the sentence could predict
(a) accent location, (b) information status, and (c) new vs. contrastive status.
Accent Location - correct trials
The overall Wilks's lambda was significant, A = .071, X2 (24) = 1102.85, p <.001,
indicating that the acoustic measures could differentiate among the three accent locations.
In addition, the residual Wilks's lambda was significant, A =.283, X2(l 1)= 536.01, p <
.001, indicating that some predictors could still differentiate accent location after
partialling out the effects of the first function. Figure 1 indicates a separation of the focus
locations on the discriminant functions.
When we tried to predict focus location, we were able to correctly classify 94.6%
of the sentences in our sample. To assess how well the classification would perform on a
new sample, we performed a leave-one-out classification and correctly classified 93.4%
of our sentences. At individual sentence locations, the discriminant function was able to
correctly classify subject focus 98% of the time, verb focus 93% of the time, and object
focus 94% of the time.
Canonical Discriminant Functions
U
U
.EIwif.
mum
Focus Location
Grao
I
Centrolds
" Ungrouped Cases
.-U
*
- omelet
- M
fried
.3i
U
-6 4
-6
" DanDn
-4
-2
0
2
4
6
Function 1
Figure1: Separationoffocus locationson two discriminantfunctions in Experiment 1
Information Status - correct trials
The overall Wilks's lambda was significant, A = .846, x2(24) = 78.37, p < .00 1,
indicating that the acoustic measures could differentiate among the information status
types. In addition, the residual Wilks's lambda was significant, A =.950, X2(11) = 24.06,
p < .05, indicating that some predictors could still differentiate information status after
partialling out the effects of the first function. Figure 2 indicates a separation of the
information status on the discriminant functions.
When we tried to predict information status, we were able to correctly classify
50.0% of the sentences in our sample. To assess how well the classification would
perform on a new sample, we performed a leave-one-out classification and correctly
classified 45.8% of our sentences. For individual levels of information status, the discrim
function was able to correctly classify new focus 54% of the time, contrastive focus 41%
of the time, and wide focus 71% of the time.
Canonical Discriminant Functions
aIF
a
Informaion Status
o] GroupCentrolds
* Contrastive
" New
. Wide
-3
-2
-1
0
1
2
3
4
Function 1
Figure2: Separation of information status on two discriminantfunctionsfor
Experiment 1.
New vs. Contrastive- correct trials
The overall Wilks's lambda was significant, A = .914, X2(12) = 37.47, p <.001,
indicating that the acoustic measures could differentiate between the new and contrastive
conditions. When we tried to predict new vs. contrastive status, we were able to correctly
classify 65.2% of the sentences in our sample. To assess how well the classification
would perform on a new sample, we performed a leave-one-out classification and
correctly classified 59.5% of our sentences.
Senance Type
Wde Focus
NowSue ct
NewVerb
NewObject
CantSuIdect
CantVerb
CantOdect
Senance Type
160
180
Senance Type
Vde Focus
FdeFocus
New~udect1601
NwSubject
.14
0
N
s
NewbSubject
No.
w Verb
wVrb
IL....
Nobwbject
V 120
2
ContSubdect
Cont Verb
Phw0Ibject
ContSubject
12
ContVerb
CCmi
2
100
Damn
fried
an
omdat
ContOlect
yesterday
2
00nnContOwect
Med
Damn
an
orndet
yesterday
Figure3: Average value offour acousticfeatures across every word and every
condition in Experiment 1. The top left graph indicates the average of the sum of the
duration of each word and the durationof anyfollowing silence, in seconds. The top
rightgraph indicatesthe average maximum intensity in decibels. The bottom left
graph indicatesthe average mean pitch, in Hertz The bottom rightgraph indicatesthe
averagemaximum pitch, in Hertz.
Results - Perception
140
120
100
Subject Choice
80
WContrastive Object
§Contrastive Ve rb
60
RContrastive Subject
40
M New Object
ENew Verb
20
[_-[New Subject
WiNde Focus
0
Wide Focus New Verb Cont Subject Cont Object
NewSubject NewObject ContVerb
Sentence Type
Figure4: Total count of Listeners' condition choice by sentence type
Wide Focus
New Verb
Cont Subje ct
C
New Subje ct
New Obje ct
Cont Verb
Sentence Type
Figure 5: Mean Listener accuracy by condition
1.00 y
.40
WideFocus
Information Status Type
New
Contrastive
1
Damon
fr ed
omd et
Focus Location
Figure 6: Mean Listener accuracy collapsed by Information Status (left) andfocus
location (right).
Listeners' choices of question sorted by the intended question are plotted in Figure 4, and
their overall accuracy percentage by condition is plotted in Figure 5. Listeners' overall
accuracy was 82%. An omnibus ANOVA on accuracy means by condition demonstrated
a significant effect of condition, such that some conditions were answered more
accurately than others, F(6) = 124.24, p <.001. Tukey post-hoc comparisons revealed
that accuracies for the Wide focus condition (51%) was significantly lower than all of the
other conditions. Individual subject accuracy ranged from 52-97%, and there were
significant differences between listeners, F(9) = 85.07, p<.001. There were no significant
differences in accuracy across items, F(13) = 1.42, p = .143.
Listeners' condition choice accuracy collapsed by information status is plotted in Figure
6 (left). Listeners accurately chose the wide focus condition 51% of the time, the new
71
focus condition 93% of the time, and the contrastive focus condition 99% of the time. An
overall ANOVA on the three different types of information status revealed significant
differences across conditions, F(2) = 1077.47, p <.001. Tukey post-hoc comparisons
revealed differences between all three conditions, such that constrastive conditions were
answered more accurately than either new or wide focus conditions, and new conditions
were answered accurately more than wide focus conditions.
Listeners' condition choice accuracy collapsed by focus location is plotted in Figure 6
(right). Listeners accurately chose the subject focus condition 93% of the time, the verb
focus condition 84% of the time, and the object focus condition 86% of the time. An
overall ANOVA on the three different types of focus location revealed significant
differences across conditions, F(2) = 34.89, p <.001. Tukey post-hoc comparisons
demonstrated that Listeners were more accurate select focus on the subject ("Damon"),
than on the other two constituents.
Discussion
The acoustic results of Experiment 1 indicate that accented words have longer durations
than their de-accented counterparts, incur larger pitch excursions, are more likely to be
followed by silence, and are produced with greater intensity. This result replicates
previous studies which showed that accents are realized by a variety of acoustic factors.
The perception results reflect the acoustic results well, in that listeners were highly
successful in discriminating between the three narrow focus conditions which focused the
Subject, Verb, and Object respectively.
The perception results also demonstrated that listeners were least successful in
discriminating wide focus ("What happened yesterday?") from the other conditions, and
confused wide focus most often with New Object focus ("What did Damon fry
yesterday?"). This result is in accordance with previous work demonstrating that an
accent on the final metrically stressed syllable of a widely focused phrase can project to
the entire phrase (Selkirk, 1995; Gussenhoven, 1999). In fact, Selkirk's account of focus
projection suggest that "Damon fried an omelet yesterday" produced with a pitch accent
only on "omelet" should be ambiguous between the new object interpretation and the
wide focus interpretation. Although listeners did confuse new object focus most often
with wide focus, they were significantly more likely to chose the wide focus condition
than the new object focus condition, suggesting that these two conditions were not
produced with the same acoustics. It may be the case that, in order to produce wide
focus, speakers were not simply producing only accent on the object.
Despite revealing multiple reliable acoustic differences between focused and nonfocused elements, the discriminant analyses performed on the productions showed only
minimal differences between focus associated with new information and focus associated
with contrastive information. Specifically, the discriminant results demonstrate only fair
classification accuracy for the information status of the focused element.
The production results, as indicated by the results of the discriminant function
analysis, suggest that Speakers disambiguated focus location with a combination of word
duration and silence, intensity, mean FO, and maximum FO. The results of the perception
study suggest that Listeners were highly accurate in determining the location of intended
focus from the Speaker's use of these acoustic cues. On the other hand, however, the
production results as indicated by the discriminant function results, suggest that the
Speakers did not successfully disambiguate information status with acoustic cues.
72
However, the perception results suggest that Listeners were able to successfully
disambiguate information status from the Speakers' productions.
The reason that the Listeners were successful in determining information status when the
discriminant function was not successful was probably that the Listeners had an
additional cue to the disambiguate contrastive focus from either new or wide focus.
Specifically, contrastive focus conditions were always begun with a 'No,' which
presumably served to disambiguate the information status without the need for acoustic
information. We tested this possibility in a second perception study on the productions
from Experiment 1.
Experiment 1A
In a second listening experiment, we spliced all of the "No"s out of the
contrastive answers so that the Listener would not have this explicit cue to the contrastive
status of the focused element. We also spliced out the beginning of the non-contrastive
answers so that any residual silence in these answers would not be a cue for the listener.
The resulting sentences then all had the form: "Damon fried an omelet yesterday."
Although we intended to have several new naive listeners do the question-choice task
with these spliced answers, we abandoned that plan after the author piloted this new task.
Her results, illustrated in Figure X, indicate that, although she demonstrated high
accuracy in deciding which constituent was being focused, she was at chance to decide
whether the accent was new or contrastive.
100
80
Subject Choice
60
Constrastive Object
E]Constrastive Verb
MContrastive Subject
40
M New Object
New Verb
20
OiNew Subject
U
Wide Focus
0
4
+_~
41
C
0-
C'
C
'6.'6
&~
Sentence Type
Figure 7: Author's condition choice by correct condition for Experiment 1A
productions with "No "'s removed.
.9-
.7
.5
.3
New Verb
Contra stive Subject Contrastive Obj ect
Wide Focus
Contra stive Ve fb
New Subje ct
New Obje ct
Figure 8: Author's accuracy by condition for Experiment 1A.
WideFocus
Information Status Type
New
Contrastive
Damon
fried
omdet
Focus Location
Figure 9: Mean author accuracy collapsed by Information Status (left) andfocus
location (right).
Results
The author's choice of question sorted by the intended question is plotted in Figure 7.
The author's overall accuracy was 48%. An omnibus ANOVA on accuracy means by
condition demonstrated a significant effect of condition, such that some conditions were
answered more accurately than others, F(6) = 5.10, p <.001. Tukey post-hoc
comparisons revealed that accuracy for the new object condition was significantly higher
than for the new verb, contrastive subject, or contrastive object conditions.
The author's condition choice accuracy collapsed by information status is plotted in
Figure 9 (left). She accurately chose the wide focus condition 58% of the time, the new
74
focus condition 56% of the time, and the contrastive focus condition 42% of the time. An
overall ANOVA on the three different types of information status revealed significant
differences across conditions, F(2) = 5.01, p <.001. Tukey post-hoc comparisons
revealed differences between all three conditions, such that constrastive conditions were
answered less accurately than either new or wide focus conditions.
The author's condition choice accuracy collapsed by focus location is plotted in Figure 9
(right). She accurately chose the subject focus condition 96% of the time, the verb focus
condition 90% of the time, and the object focus condition 92% of the time. An overall
ANOVA on the three different types of focus location revealed no significant difference
across conditions, F(2) = 1.93, p =.15.
Discussion
The perception results of Experiment la suggest that Listeners' high accuracy in selecting
the correct information status conditions in Experiment 1 was probably due to the
presence of the disambiguating 'No,' rather than to any cues in the prosody of the
speaker.
The lack of a strong difference between information status conditions could have
resulted from one of two reasons. First, it could be the case that speakers do not
acoustically differentiate between new and contrastive accents. Alternatively, however, it
could mean that speakers do not provide prosodic cues to information that is predictable
from the context, as contrastive status was in this experiment, from the inclusion of the
'No' in the answer. A growing body of literature suggests that speakers are less likely to
produce intonational cues for sentence structures with more predictable meanings from
the contexts (Snedeker & Trueswell, 2004). We tested this possibility in Experiment 2,
where we specifically removed the additional cue to the discourse status of the focused
constituent ("No").
Experiment 2
Method
Participants
We recorded 14 pairs of subjects for this study. Each subject received $10/hour
for his/her participation. Four pairs of subjects were excluded for poor performance, in
that the Listeners in the pair chose the correct sentence 20% of the time or less.
Materials
The materials for Experiment Two were identical to those from Experiment 1 described
above with the exception that the word "No" was excluded from the contrastive
conditions and the words "Iheard that" were added to each of the seven conditions. An
example item is presented in Table 3.
Condition
Status
Focused
Argument
1
New
wide
2
New
S
3
New
V
SetupQuestion
What happened yesterday?
Who fried an omelet yesterday?
What did Damon do to an omelet
yesterday?
Target
I heard that Damon fried
an omelet yesterday.
I heard that Damon fried
an omelet yesterday.
I heard that Damon fried
an omelet yesterday.
4
New
0
What did Damon fry yesterday?
5
Contrastive
S
6
Contrastive
V
7
Contrastive
0
Did Harry fry an omelet yesterday?
Did Damon bake an omelet
yesterday?
Did Damon fry a chicken
yesterday?
I heard that Damon fried
an omelet yesterday.
I heard that Damon fried
an omelet yesterday.
I heard that Damon fried
an omelet yesterday.
I heard that Damon fried
an omelet yesterday..
Table 3: Example item from Experiment 2
Procedure
The procedure for Experiment 2 was the same as that described for Experiment 1.
Results - Production
We tested the acoustic features we had identified in Experiment 1 on the new productions
in Experiment 2. Once again, we conducted three discriminant analyses to determine
whether the measures of (1) duration + silence, (2) maximum pitch, (3) mean pitch, and
(4) maximum intensity on the three critical words in the sentence could predict (a) accent
location, (b) information status, and (c) new vs. contrastive conditions.
Accent Location - correcttrials
The overall Wilks's lambda was significant, A = .061, X2(24) = 1480.62, p <.001,
indicating that the acoustic measures could differentiate among the three accent locations.
In addition, the residual Wilks's lambda was significant, A =.279, X2(1 1) = 674.66, p <
.001, indicating that some predictors could still differentiate accent location after
partialling out the effects of the first function. Figure 10 indicates a separation of the
focus locations on the discriminant functions.
When we tried to predict focus location, we were able to correctly classify 97.0%
of the sentences in our sample. To assess how well the classification would perform on a
new sample, we performed a leave-one-out classification and correctly classified 96.5%
of our sentences. At individual sentence locations, the discriminant function was able to
correctly classify subject focus 93% of the time, verb focus 94% of the time, and object
focus 87% of the time.
Canonical Discriminant Functions
8
6
0
.
01
Focus Location
a
".
0.
-2"
Group Centroids
0Ungroupe d Case s
4
om elet
Function 1
Figure 10: Separationoffocus locationson two discriminantfunctions
Information Status -correct trials
The overall Wilks's lambda was significant, A =.705, X2(24) = 216.19, p <.001,
indicating that
the acoustic measures could differentiate among the three information
status conditions. In addition, the residual Wilks's lambda was significant, A = .963,
X2(11) = 23.08, p < .05, indicating that some predictors could still differentiate
information status after partialling out the effects of the first function. Figure 11I
indicates a separation of the information status on the discriminant functions.
When we tried to predict information status, we were able to correctly classify
50.9% of the sentences in our sample. To assess how well the classification would
perform on a new sample, we performed a leave-one-out classification and correctly
classified 48.3% of our sentences. For individual levels of information status, the
discriminant function was able to correctly classify new focus 52% of the time,
contrastive focus 44% of the time, and wide focus 70% of the time.
Canonical Discriminant Functions
6
4 1
.
0
*
*
01
Information Status
0 GroupCentroids
" Contrastive
*
L
New
-4
" Wide
-4
-3
-2
-1
0
1
2
3
4
Function 1
Figure 11: Separation of information status types on two discriminantfunctions
New vs. Contrastive- correct trials
The overall Wilks's lambda was significant, A =.908, X2(24) = 51.17, p <.001,
indicating that the acoustic measures could differentiate among new and contrastive
conditions. When we tried to predict new vs. contrastive information status, we were
able to correctly classify 59.8% of the sentences in our sample. To assess how well the
classification would perform on a new sample, we performed a leave-one-out
classification and correctly classified 57.7% of our sentences.
Information Status - including "I"
We also wanted to determine how important the prosody of the "I heard that" was
to the differentiation of the information status of the sentences in Experiment 2. To
investigate this question, we performed a stepwise discriminant function analysis which
included as predictors the four acoustic factors we had identified in Experiment 1
(duration + silence, mean pitch, maximum pitch, maximum intensity) for the subject
("Damon"), verb ("fried"), and object ("omelet"), as well as for the first three words of
the sentence ("I heard that"). The analysis revealed that the most important variables for
the differentiation of information status were (1) the duration + silence of "I",(2) the
maximum pitch of "I", and (3) the Maximum intensity of "I". Following that, we
conducted another analysis in which we included as predictors the duration + silence,
mean pitch, maximum pitch, and maximum intensity of the subject, verb, object, and "I."
The overall Wilks's lambda was significant, A = .469, X2(32) = 454.79, p < .001,
indicating that the acoustic measures could differentiate among the three information
status conditions. In addition, the residual Wilks's lambda was significant, A =.815,
X2(15) = 122.76, p <.001, indicating that some predictors could still differentiate
information status after partialling out the effects of the first function. Figure 12
indicates a separation of the information status on the discriminant functions.
When we tried to predict information status, we were able to correctly classify
69.2% of the sentences in our sample. To assess how well the classification would
perform on a new sample, we performed a leave-one-out classification and correctly
classified 67.4% of our sentences. For individual levels of information status, the
discriminant function was able to correctly classify new focus 72% of the time,
contrastive focus 64% of the time, and wide focus 75% of the time.
Canonical Discriminant Functions
43
2
0
0
U
Mr,
U
01
-1
A
Eu
-21
*0
a
Mu
p.
Information Status
o3 Group Centroids
sos
* Contrastive
-3
New
" Wide
-4
-4
-3
-2
-1
0
1
2
3
Function 1
Figure 12: Separation of information status types on two discriminantfunctions
New vs. Contrastive- including "I"
The overall Wilks's lambda was significant, , A = .606, X2 (16) = 258.37, p <.001,
indicating that the acoustic measures could differentiate among new and contrastive
conditions. When we tried to predict new vs. contrastive conditions, we were able to
correctly classify 79.3% of the sentences in our sample. To assess how well the
classification would perform on a new sample, we performed a leave-one-out
classification and correctly classified 78.9% of our sentences. This classification
accuracy is much higher than that achieved using the acoustics of the subject, verb, and
object alone, indicating that speakers were primarily using the prosody of "I"to encode
the difference between information status type.
Sentence Type
66
MAdeFocus
New
Sulbect
New~r
NewOect
C
E
E 62
Contestfe Subject
as
Contustve Wib
2 s
Contreeste
Object
SentenceType
SentenceType
V deFocus
VAdeFocus
NewSubject
oNewSubject
NowwVrb
NewVrb
NewCf*ct
NewOect
Contrasthe
Subject
Contrastive
Subject
Contrastve
W rb
Contrastive %Arb
ContrestieObjct
heard
tat
fried
Damon
an
omdet
yesterday
heard
Damon
an
Contrasthe Object
ordet
yesterday
Figure13: Average value offour acousticfeatures across every word and every
condition in Experiment 2. The top left graph indicates the average of the sum of the
duration of each word and the duration of anyfollowing silence, in seconds. The top
rightgraph indicatesthe average maximum intensity in decibels. The bottom left
graph indicates the average mean pitch, in Hertz. The bottom rightgraph indicates the
average maximum pitch, in Hertz.
80
Results - Perception
140
120 r .
Subject Choice
1001
WContrastive Object
801
OContrastive Verb
60
6Contrastive Subject
New Object
40lM
ENew Verb
4-
201
U
0
cDNe w Subject
0-
C%
0
Wi de Focus
Sentence Ty pe
Figure14: Total count of Listeners condition choice by sentence type
1.0'
.8
AI
O~
6
2
.4
Wide Focus
Ne w Ver b
Contra stive Subject Contrastive Object
Contra stive Verib
New Obje ct
Ne w Subje ct
Sentence Type
Figure15: Mean Listener accuracy by condition
Figure 16: Mean Listener accuracy collapsed by Information Status (left) andfocus
location (right).
Listeners' choices of question sorted by the intended question are plotted in Figure 14,
and their overall accuracy percentage by condition is plotted in Figure 15. Listeners'
overall accuracy was 70%. An omnibus ANOVA on accuracy means by condition
demonstrated a significant effect of condition, such that some conditions were answered
more accurately than others, F(6) = 8.39, p < .001Individual subject accuracy ranged
from 58-80%, and there were significant differences between listeners, F(9) = 26.00,
p<.001. There were no significant differences in accuracy across items, F(13) = 1.1.
Listeners' condition choice accuracy collapsed by information status is plotted in Figure
16 (left). Listeners accurately chose the wide focus condition 73% of the time, the new
focus condition 82% of the time, and the contrastive focus condition 64% of the time. An
overall ANOVA on the three different types of information status revealed significant
differences across conditions, F(2) = 154.39, p <.001. Tukey post-hoc comparisons
revealed differences between all three conditions, such that constrastive conditions were
answered more accurately than either new or wide focus conditions, and new conditions
were answered accurately more than wide focus conditions.
Listeners' condition choice accuracy collapsed by focus location is plotted in Figure 16
(right). Listeners accurately chose the subject focus condition 93% of the time, the verb
focus condition 94% of the time, and the object focus condition 87% of the time. An
overall ANOVA on the three different types of focus location revealed significant
differences across conditions, F(2) = 37.53, p <.001. Tukey post-hoc comparisons
demonstrated that Listeners were more accurate to select focus on the subject ("Damon"),
or the verb ("fried") than on the object ("omelet").
Discussion
The acoustic results of Experiment 2 performed only on the subject, verb, and
object, once again demonstrated that speakers were consistent and successful in
disambiguating the intended position of focus. Similar to Experiment 1, focused words
were indicated with longer durations, a greater probability of post-word silence, higher
intensity, higher mean FO, and higher maximum FO. Also, as demonstrated in
Experiment 1, speakers did not systematically disambiguate information status with their
productions, as evidenced by the discrim results of classification of only 50%.
The acoustic results of Experiment 2 performed on the subject, verb, object, and
"I" indicated more systematic disambiguation of information status. Specifically,
contrastive conditions were indicated by "I"'s produced with longer durations, higher
intensity, and higher mean FO and maximum FO.
The individual graphs of single acoustic features presented in Figure 13 are
similar to those from Experiment 1, in that they demonstrate higher values of duration +
silence, FO, and intensity on focused constituents than non-focused constituents. These
results are different from those in Experiment 1,however, in the realization of FO for the
wide focus conditions. Specifically, the FO values are higher for wide focus conditions
compared to the other conditions.
Once again, the perception results demonstrated that Listeners were very accurate
in determining which constituent of the sentence was focused. They were, however, not
as accurate at determining the information status of the sentence as they were in the first
study. This second result is to be expected since the Listeners no longer had the explicit
cue of the "No" in the contrastive conditions to signal the difference between new and
contrastive meanings.
Although we designed Experiment Two with the same words in each condition in
order to encourage speakers to disambiguate between the new and contrastive readings of
the focused constituents with the accent they placed on those constituents, it was also
possible for speakers to disambiguate the discourse status of the focused constituent with
how they prosodified the first three words of the sentence (i.e. "I heard that"). In fact, a
discriminant function analysis demonstrated that the prosody of the first word ("I") of
each sentence was the strongest predictors of the difference between new and contrastive
information status.
Why might speakers have disambiguated information status most strongly with
of "Iheard that"? One possibility is that emphasizing the "I," which the
production
their
discriminant results suggest that the speakers did, serves to signal pragmatically that the
speaker means to contrast the information in the sentence that follows "I heard that" with
what his/her questioner assumes. Whether or not this is a something the speakers would
normally do to preface contrastive information is an open question, however, given that
their task was explicitly to induce their listeners to correctly choose the new or
contrastive condition.
The results of Experiment 2 once again do not provide evidence as to whether the
lack of a difference between the acoustic realization of information status is a result of a
general lack of such disambiguation in normal speech or whether, similar to Experiment
1, speakers did not disambiguate information status on the focused words because they
did so on the "I heard that." This latter possibility was explored in Experiment 3, where
speakers had only the sentence "Damon fried an omelet yesterday" available to realize
acoustic differences between information statuses.
Experiment 3
Experiment Three was designed to be the strongest test of speakers' ability to
disambiguate information status with prosody. That is, given the sentences they are
required to speak, speakers may only disambiguate the discourse status of sentence
constituents by producing different accents on those constituents. If speakers are truly
able to disambiguate between "new" and "contrastive" information with different
accents, we will see that reflected in their productions in Experiment 3.
83
Method
Participants
We recorded 17 pairs of subjects for this experiment. As before, we excluded speaker and
listener pairs in which the Listener did not achieve accuracy greater than 20%. This
resulted in the exclusion of two pairs. We also excluded one pair in which the speaker
was not a native speaker of English. Finally, we excluded two pairs of subjects who
produced unnatural prosody to effect a disambiguation between new and contrastive
conditions. Specifically, they produced contrastive accents with unnaturally emphatic
accents. These exclusions resulted in a total of 13 subjects whose productions and
perceptions were analyzed.
Materials
The materials for Experiment 3 are identical to those from Experiment 2 described above
save for the exclusion of "I heard that" from all conditions. An example item is
presented in Table 4.
Condition
Status
Focused
Argument
SetupQuestion
What happened yesterday?
1
New
wide
2
New
S
3
New
V
Who fried an omelet yesterday?
What did Damon do to an omelet
yesterday?
4
New
0
What did Damon fry yesterday?
5
Contrastive
S
Did Harry fry an omelet yesterday?
6
Contrastive
V
Did Damon bake an omelet yesterday?
7
Contrastive
0
Did Damon fry a chicken yesterday?
Target
Damon fried an
omelet yesterday.
Damon fried an
omelet yesterday.
Damon fried an
omelet yesterday.
Damon fried an
omelet yesterday.
Damon fried an
omelet yesterday.
Damon fried an
omelet yesterday.
Damon fried an
omelet yesterday..
Table 4: Example item from Experiment 3.
Procedure
The procedure for Experiment 3 was the same as that described for Experiment 1.
Results - Production
We tested the acoustic-features we had identified in Experiment 1 on the new productions
in Experiment 3. Once again, we conducted three discriminant analyses to determine
whether the measures of (1) duration + silence, (2) maximum pitch, (3) mean pitch, and
(4) maximum intensity on the three critical words in the sentence could predict (a) accent
location, (b) information status, and (c) new vs. contrastive conditions.
Accent Location - correct trials
The overall Wilks's lambda was significant, A = .095, X2 (24) = 1274.05, p <.001,
indicating that the acoustic measures could differentiate among the three accent locations.
In addition, the residual Wilks's lambda was significant, A = .329, X2 (11) = 602.50, p <
.001, indicating that some predictors could still differentiate accent location after
partialling out the effects of the first function. Figure X indicates a separation of the
focus locations on the discriminant functions.
84
When we tried to predict focus location, we were able to correctly classify 92.5%
of the sentences in our sample. To assess how well the classification would perform on a
new sample, we performed a leave-one-out classification and correctly classified 92.0%
of our sentences.
Canonical Discriminant Functions
8
61
M
Focus Location
y
-4
.
gGrot4
Centroids
Ungroupe d Case s
m.""
4"omeet
S-6
fr ied
.
L -8
Damon
____________________
-4
-2
0
2
4
6
8
Function 1
Figure17: Separationoffocus locations on two discriminantfunctions
Information Status - correct trials
The overall Wilks's lambda was significant, A = .759, X(24) = 178.29, p < .001,
indicating that the acoustic measures could differentiate among the three information
status conditions. In addition, the residual Wilks's lambda was significant, A = .947,
X2 (11)= 35.35, p <.001, indicating that some predictors could still differentiate
information status after partialling out the effects of the first function. Figure X indicates
a separation of information status on the discriminant functions.
When we tried to predict information status, we were able to correctly classify
52.0% of the sentences in our sample. To assess how well the classification would
perform on a new sample, we performed a leave-one-out classification and correctly
classified 50.0% of our sentences.
Canonical Discriminant Functions
a
.
U
U
Information Status
*=
U
-1 1
G
o
."U
GroupCentroids
" Contrastive
-2
-3 2
1U
-3
1
2
3
New
" Wide Focus
-4
Function 1
Figure 18: Separationof information status types on two discriminantfunctions
New vs. Contrastive- correct trials
The overall Wilks's lambda was significant, A =.892, X2 (24) = 62.08, p <.001,
indicating that the acoustic measures could differentiate among new and contrastive
conditions. When we tried to predict new vs. contrastive information status, we were
able to correctly classify 64.2% of the sentences in our sample. To assess how well the
classification would perform on a new sample, we performed a leave-one-out
classification and correctly classified 62.4% of our sentences.
Condition
ideFocus
WAde
Focus
NewSdect
NewSAect
New rb
New rb
NewGect
NewOect
Contas swed
contase Su
CoetMesse
W&b
Contasdve Web
Cotas66 Otied
contasiVe Oedeco
dec
Condition
WideFocus
Wde Focus
NewSie ct
NewSleWct
New
NowVrb
rb
NewOaect
NewOect
ContMs" sowed
ContasMie Sutged
Contasie Wib
Contasie Wrb
Contmse Otied
contase
Owed
Figure19: Average value offour acousticfeatures across every word and every
condition in Experiment 3. The top left graph indicates the average of the sum of the
duration of each word andthe durationof any following silence, in seconds. The top
right graphindicates the average maximum intensity in decibels. The bottom left
graph indicatesthe averagemean pitch, in Hertz. The bottom rightgraph indicatesthe
average maximum pitch, in Hert.
87
Results - Perception
200
Subject Sent Choice
Contrastive Object
ElContrastive Verb
100,
&Contrastive
*New
Subj ect
Object
ENewVerb
New Subject
0
U
0 , _o
Wide Focus
.
*c 4%
-
0-
CS
Condition
Figure20: Total count of Listeners' condition choice by sentence type
1.0
.8r
Wide Focus
New Ver b
Contra stive Subject Contrastive Object
Ne w Subje ct
New Obje ct
Contra stive Ve rb
Condition
Figure21: Mean Listener accuracy by condition
88
1.00
1.0
80
C
iN
0
ct
o.6
.4.0
.4
WideFocus
Information Status Type
L4O
New
Contra
stiveam
frled
omde
Focus Location
Figure 22: Mean Listener accuracy collapsed by Information Status (left) andfocus
location (right).
Listeners' choices of question sorted by the intended question are plotted in Figure 20,
and their overall accuracy percentage by condition is plotted in Figure 21. Listeners'
overall accuracy was 55%. An omnibus ANOVA on accuracy means by condition
demonstrated a significant effect of condition, such that some conditions were answered
more accurately than others, F(6) = 4.90, p < .001. Tukey post-hoc comparisons revealed
that Listener accuracies for the three contrastive conditions were all lower than the the
New subject condition; no other comparisons reached significance. Individual subject
accuracy ranged from 39-96%, and there were significant differences between listeners,
F(9) = 9.9, p<.001. There were no significant differences in accuracy across items, F(13)
< 1.
Listeners' condition choice accuracy collapsed by information status is plotted in Figure
22 (left). Listeners accurately chose the wide focus condition 60% of the time, the new
focus condition 70% of the time, and the contrastive focus condition 48% of the time. An
overall ANOVA on the three different types of information status revealed significant
differences across conditions, F(2) = 157.82 p <.001. Tukey post-hoc comparisons
revealed differences between all three conditions, such that constrastive conditions were
answered more accurately than either new or wide focus conditions, and new conditions
were answered accurately more than wide focus conditions.
Listeners' condition choice accuracy collapsed by focus location is plotted in Figure 22
(right). Listeners accurately chose the subject focus condition 88% of the time, the verb
focus condition 86% of the time, and the object focus condition 80% of the time. An
overall ANOVA on the three different types of focus location revealed significant
differences across conditions, F(2) = 33.49, p <.001. Tukey post-hoc comparisons
demonstrated that Listeners were more accurate to select focus on the subject ("Damon"),
or the verb ("fried") than on the object ("omelet").
Discussion
Focus location
In the third experiment, we observed the same high accuracy in determining the
focus location of the accent in the sentence by both listeners and by a discriminant
function analysis using duration + silence, maximum pitch, mean pitch, and maximum
intensity. These results suggest, once again, that speakers are consistently indicating
89
focus location using this set of acoustic features, and that listeners are able to interpret
these cues as indicating the location of sentence focus. Moreover, these results suggest
that different speakers are using the same set of features across multiple experiments to
indicate focus location.
It must be noted, however, that the task used in each of the first three experiments
was not a natural one. By making the speaker's goal to induce his/her listener to choose
the correct question, speakers may have been producing acoustic disambiguations that
they would not normally produce. In Experiment 4, we investigated whether speakers
would still disambiguate focus location with the same acoustic features even when they
were not aware of a need to explicitly disambiguate this information for their listener.
Information status
The results of Experiment 3 once again demonstrate that speakers were not
systematically disambiguating information status with their productions. In fact, across
Experiments 1, 2, and 3, speakers overall did not produce acoustic cues to information
status disambiguation which allowed the discriminant function analysis to classify
information status with better than 50-52% accuracy across all experiments. These
results suggest that speakers did not alter the way they produced the subject, verb, and
object of the sentence whether or not there were other ways to disambiguate the sentence
(e.g. with "No" in Experiment 1 or with "I heard that" in Experiment 2).
These results, however, do not lead to the conclusion that speakers do not
acoustically disambiguate information status. As previously mentioned, the task may not
have elicited speakers' normal production behavior. Specifically, the fact that the
speakers' task was to induce his/her listener to choose the correct question, speakers may
have settled on a strategy-any strategy-that would allow their listeners to successfully
choose the right question. Different speakers may have pursued very different strategies
which, although allowing their individual listener to perform accurately, did not overall
result in systematic acoustic disambiguation of information status.
We investigated this possibility by performing a discriminant function analysis on
the two best Speakers from Experiment 3. The best Speakers were those whose Listeners
performed with the highest accuracy, averaging 83% accuracy. The classification
accuracy achieved by the discriminant analysis is presented in Figure 21. The
discriminant function analysis was able to correctly classify new focus 76% of the time,
contrastive focus 90% of the time, and wide focus 89% of the time. The associated
Listener classification accuracy is presented in Figure 22. Listeners of the two best
Speaker were able to correctly classify new focus 77% of the time, contrastive focus 95%
of the time, and wide focus 64% of the time.
The results of analyses performed on the best speakers from Experiment 3 suggest
that, under certain circumstances, speakers do systematically disambiguate information
status acoustically. It may be the case that this disambiguation is not consistent across
speakers, and so results collapsed over multiple speakers washes out the disambiguating
cues utilized by that individual speakers.
Canonical Discriminant Functions
4'
3'
0 13
2.
A- t
'a
M
U
.
1I
0
Informaion Status
W'he momE
ann
o
GroupCentrolds
* Contrastive
New
" Wide Focus
-4
-2
0
2
4
6
Function 1
Figure23: Separation of informationstatusproductionsof the two best speakersfrom
Experiment 3 on two discriminantfunctions
New
Contrastiv
Figure24: Listener accuracy on information status conditionsfrom the best two
Speakersfrom Experiment 3.
Experiment 4
The first three experiments that we conducted are valuable in that they constitute several
major methodological advantages over previous work investigating the production and
comprehension of accents in English. First, they constitute one of the first attempts to
investigate whether and how naive speakers disambiguate focus and discourse status of
91
sentence constituents, as well as whether naYve listeners are sensitive to such
disambiguation. Second, they represent one of the first systematic evaluations of the
acoustic features of English accents produced by naive speakers.
It is important to note, however, that the productions from speakers in the first
three experiments are not naturally, spontaneously produced. In order to maximize the
amount of data we could gather from a small number of subjects, we chose to utilize a
within-subjects design and to provide the words that we intended the speakers to produce.
Both of these decisions could be argued to take away from the applicability of the results.
Specifically, the fact that speakers were aware of the seven possible interpretations of
each sentence, and of their goal of providing their listener with means to chose the correct
question, encouraged them to maximally differentiate their productions to the best of
their ability. Furthermore, the speakers always knew the words they were going to
produce before they produced them.
To account for these limitations of the first three experiments, the fourth
experiment in this sequence was designed to investigate accent production in more
naturalistic productions. To the extent that the speakers' behavior in Experiment 4
mimics that of the speakers in Experiments 1-3, we can consider that behavior to be
spontaneous, and not the result of unnatural communication pressures we placed on our
subjects.
Method
Participants
We recorded seven pairs of subjects in Experiment 4, for a total of 14 speakers. Two
speakers had to be replaced, because they failed to produce the correct words on more
than 25% of trials. Subjects were paid $10/hour for their participation.
Materials
Condition
Status
Focused
Argument
SetupQuestion
1
New
wide
2
New
S
Who fried an omelet last night?
3
New
V
What did Lena do to an omelet last night?
4
New
0
What did Lena fry last night?
5
Contrastive
S
Did Harry fry an omelet last night?
6
Contrastive
V
Did Lena bake an omelet last night?
7
Contrastive
0
Did Lena fry a chicken last night?
What happened last night?
Target
Lena fried
last night.
Lena fried
last night.
Lena fried
last night.
Lena fried
last night.
Lena fried
last night.
Lena fried
last night.
Lena fried
last night.
an omelet
an omelet
an omelet
an omelet
an omelet
an omelet
an omelet
Table 5: Example Item from Experiment 4.
Procedure
The experiment was conducted in two parts. The first part was a training session,
where participants learned the correct names for pictures of people, actions, and objects.
In the second part of the experiment the two participants took turns serving as the
Questioner and the Speaker. The Questioner was produced question for the Speaker, and
then the Speaker produced the answer to the question which was indicated by pictures
92
presented on his/her screen. The next sections will describe each portion of the
experiment in more detail.
PictureNaming Training
In a preliminary training session, both participants learned the mapping between
96 pictures and names, so that they could produce the names from memory during the
main experiment. In a power point presentation, each picture, corresponding to a person,
an action, or an object, was presented with its intended name. In order to facilitate
memorization, the pictures were presented in alphabetical order. Participants were
instructed to go through the power point at their own pace, with the goal of learning the
mappings.
When they felt they had learned the mappings, participants waere given a picturenaming test, which consisted of 27 items from the full list of 96. Participants were told of
their mistakes, and, if they incorrectly named four or more of the items on the test, they
were instructed to go back through the power point to improve their memory of the
picture-name mappings.
Question-Answer Experiment
After both members of the pair had successfully learned the picture-name
mappings, the members were randomly assigned to the Questioner and Speaker roles.
They sat at computers in the same room such that neither could see the other's screen.
The Questioner saw a question on his/her screen which he produced aloud for the
Speaker. The Speaker then used the pictures on his/her screen to answer the Questioner.
93
Results -Production
66 '
.4
64
Condition
Condition
.3deFocus
62
+Ne*rb
+
E 60
-
.2
NewVu-b
N6wGwct
New
Contrastie Subject
.1
Damon
2
deFocus
NewSuloct
t
NawSuft
fried
an
anda
last
Contnsth Subject
Contrastive Verb
C
Contrastw %Arb
Contrasthe
Object
2
night
56
Damon
210
240
200
230
190
220
180
Condition
ContrasthA
Obect
fried
an
amde
last
210
night
Condition
Ade Focus
V deFocus
170
200
~160
Ne-W
190o
--
NewSbject
NewSbject
~
E
EU
ContstVeWrb
2
age ct
a 58
140
_
Damon
Object
_Contastie
fried
an
on det
last
night
Conteste Wrb
2
170
Damsn fried
Contastive Object
an
aet
last
night
Figure25: Average value offour acousticfeatures across every word and every
condition in Experiment 3. The top left graphindicates the average of the sum of the
duration of each word and the durationof any following silence, in seconds. The top
rightgraph indicatesthe average maximum intensity in decibels. The bottom left
graph indicatesthe averagemean pitch, in Hertz. The bottom rightgraph indicatesthe
average maximum pitch, in Hertz.
As in the previous experiments, three discriminant function analyses were conducted to
determine whether the measures of (1) duration + silence, (2) maximum pitch, (3) mean
pitch, and (4) maximum intensity on the three critical words in the sentence could predict
(a) accent location, (b) information status, and (c) new vs. contrastive.
Accent Location
The overall Wilks's lambda was significant, A =.67, X2 (24) = 165.44, p <.001,
indicating that the acoustic measures could differentiated among the three accent
locations. In addition, the residual Wilks's lambda was significant, A = .92, x2 (G1)
35.76, p <.001, indicating that some predictors could still differentiate accent location
after partialling out the effects of the first function. Figure X indicates a separation of the
focus locations on the discriminant functions.
When we tried to predict focus location, we were able to correctly classify 60% of
the sentences in our sample. The kappa value of .40 indicated moderate accuracy in
classification performance. To assess how well the classification would perform on a
94
new sample, we performed a leave-one-out classification and correctly classified 58% of
our sentences. At individual sentence locations, the discrim function was able to correctly
classify subject focus 69% of the time, verb focus 51% of the time, and object focus 60%
of the time.
Canonical Discriminant Functions
4
U
U
2
0
.
~Group~
*
Centroids
a
-4
om elet
M
LL
" Ungroupe d Case s
i
Efried
-6
Darnon
"___________________
-4
-2
0
2
4
6
Function 1
Figure26: Separationoffocus locations on two discriminantfunctions in Experiment
4.
Information Status
2 (24) = 37.47, p < .05, indicating
The overall Wilks's Lambda was significant, A = .93, x
that the acoustic features could discriminate between the three information status
conditions. Figure 27 indicates the separation of information status groups on the two
discriminant functions.
When we tried to classify information status, we were able to correctly classify
42% of cases. The kappa value of -.11, however, indicated that agreement was lower
than what would be expected by chance, and therefore not reliable. In fact, when we
estimated the percent of new cases that would be correctly classified, using a leave-oneout technique, we correctly classified only 36% of cases. For individual levels of
information status, the discrim function was able to correctly classify new focus 52% of
the time, contrastive focus 27% of the time, and wide focus 60% of the time.
Canonical Discriminant Functions
EU.
1P
U
koft~
af *u
*
Uetrid
a.
0
otraspCetid
-2
Ne w
*Wi de
Function 1
Figure 27: Separation of information status on two discriminantfunctions in
Experiment 4.
New vs. Contrastive
The overall Wilks's Lambda was not significant, A = .98, ) (12) = 10.61, p = .56,
indicating that the acoustic features selected could not discriminate between new and
contrastive sentences.
Discussion
Focus Location
As in the previous experiments, speakers did consistently provide acoustic cues
which served to disambiguate focus location for their listeners. In addition, they
indicated focus with the same cues they had used in previous experiments: increased
duration, higher intensity, higher mean FO, and higher maximum FO. These results are
noteworthy in that, unlike in the previous experiments, the speakers in the current
experiment were not explicitly trying to induce their listeners to choose the correct
meaning. Therefore, these results suggest that the acoustic cues we identified are those
which are normally used by speakers in natural conversation to indicate the focused
material in a sentence.
InformationStatus
The results from Experiment 4 indicate, once again, that speakers do not
systematically differentiate between information status conditions. Specifically, a
96
discriminant function analysis could not classify speakers' productions by information
status any better than chance would predict. There are at least two possible explanations
for this result. The first, as previously suggested, is that speakers do not systematically
differentiate information status with prosody. As previously stated, it could be the case
that different speakers have different ways of encoding information status with their
prosody that, when averaged with the prosody of other speakers, can no longer
effectively disambiguate between conditions. The second possible explanation is that the
task did not mimic natural communication pressures that would induce speakers to
produce differences in prosody.
It should be noted that the data from Experiment 4 were not pre-screened in the
same way as those productions from the first three experiments. Specifically, because
there was no listener engaged in a meaning task in Experiment 4, there is no way to select
(a) only those speakers who were engaged in the task and were prosodically
differentiating their productions, or (b) those individual productions which bear meaning.
Future analyses of these data should be performed on only those trials which are shown
to convey information about focus location or information status to a new set of
Listeners.
General Discussion
The results from the four experiments presented in this chapter represent
important contributions to basic research on the information that is conveyed by prosody.
Specifically, these data demonstrate that speakers do systematically provide cues to the
location of focused material with prosody, and that they use a combination of duration,
intensity, and pitch indicate this information. Furthermore, speakers provide cues to
focus location whether or not the task explicitly demands it.
The second important result from this series of experiments is that that speakers
do not systematically provide cues to information status with prosody. That is, speakers
do not indicate the new or contrastive status of focused words consistently. In addition, it
appears that speakers are not any more or less likely to provide acoustic cues to
information status when the context in which the information occurs contains additional
cues to information status. Specifically, whether or not the contrastive sentences were
preceded with a "No" did not influence the overall usefulness of acoustic cues to
contrastive focus.
In addition, these experiments represent a significant improvement over previous
investigations of the relationship between prosody and meaning for the following
reasons: First, we did not exclude speakers based on our perceptions of their productions.
Speakers who were excluded were either not providing information to their Listeners, or
were obviously not taking the task seriously. Second, we investigated the acoustic
realization of different sentence positions, and multiple items, to ensure that differences
we observed were not limited to certain syntactic positions or to a limited set of materials.
Third, we utilized multiple, untrained speakers to ensure that our results are
generalizeable to all speakers and are not due to speakers prior beliefs about what type of
accent signals a particular type of information. Finally, we elicited and selected for
analysis productions using a meaning task, rather than basing differences on perceptual
differentiability or on ratings of the appropriateness of certain prosodic contours for
particular purposes.
97
These studies open the door to future investigations of the ways in which
information status is realized in speech and to whether different accent categories indicate
any differences in meaning. Are there any circumstances under which speakers will
produce systematic cues to new and contrastive status? If so, are these different
information types realized with consistent acoustic cues, or consistent intonational
shapes? If, on the other hand, there are no consistent differences between accents that
mark new and contrastive information status, is there support for any categories of
accents, or do accents vary continuously depending on multiple discourse, speaker, and
context factors? Accents may be only one of many cues to information status that cannot
be interpreted in the absence of their full linguistic context.
References
Bartels, C., Kingston, J., (1994). Salient Pitch Cues in the Perception of Contrastive
Focus. The Journalofthe Acoustical Society ofAmerica, Volume 95, Issue 5, 2973-
Beaver, D., Clark, B., Flemming, E., Jaeger, T.F., & Maria Wolters, 2007. When
Semantics meets Phonetics: Acoustical studies of second occurrence focus. Language,
83(2), 245-276.
Beckman, M. and Hirschberg, J. (1994) The ToBI annotation conventions. Technical
report, Ohio State University, Columbus.
Beckman, M. and Pierrehumbert, J. (1986) Intonational structure in Japanese and
English. Phonology Yearbook, 3, 255-309.
Beckman, M. E., Hirschberg, J., & Shattuck-Hufnagel, S. (2005). The original ToBI
system and the evolution of the ToBI framework. In S.-A. Jun (ed.) Prosodic Typology
The Phonology of Intonation and Phrasing, Oxford University Press, Chapter 2: 9-54.
Beckman, M., and Ayers-Elam, G. (1994) Guidelines for ToBI Labelling. v. 2. Ohio
State University. Available at www.ling.ohio-state.edu/research/ phonetics/EToBI/.
--
Birch, S. and Clifton, C. Jr. (1995) Focus, accent, and argument structure: effects on
language comprehension. Language and Speech, 38 (4), 365-391.
Bock & Mazzella (1983) Intonational marking of given and new information: Some
consequences for comprehension. Memory and Cognition, 11, 64-76.
Boersma, Paul & Weenink, David (2006). Praat: doing phonetics by computer (Version
4.3.10) [Computer program]. Retrieved June 3, 2005, from http://www.praat.org/
Bolinger, D. (1958). A Theory of Pitch Accent in English. Word 14, 109-149.
Bolinger, D. (1961). Contrastive accent and contrastive stress. Language, 37, 83-96.
Brugos, A., Shattuck-Hufnagel, S, & Veilleux, N. (2006) Transcribing Prosodic Structure
of Spoken Utterances with ToBI, January (IAP) 2006, chapter 2.12
http://ocw.mit.edu/OcwWeb/Electrical-Engineering-and-Computer-Science/691 lJanuary--IAP--2006/CourseHome/index.htm
Calhoun, S. (2004). Phonetic dimensions of intonational categories: the case of L+H* and
H*. In Proceedingsof the InternationalConference on Spoken LanguageProcessing,
Nara: Japan.
Calhoun, S. (2005). It's the difference that matters: An argument for contextuallygrounded acoustic intonational phonology. In Linguistics Society ofAmerica Annual
Meeting, Oakland, California.
Carlson, K., Clifton, C., Jr., & Frazier, L. (2001). Prosodic boundaries in adjunct
attachment. JournalofMemory and Language, 45, 58-81.
Choi, J., Hasegawa-Johnson, M., & Cole, J. (2005). Finding intonational boundaries
using acoustic cues related to the voice source. JournalofAcoustical Society ofAmerica,
188 (4), 2579-2587.
Cooper, W., Eady, S. & Mueller, P. (1985) Acoustical aspects of contrastive stress in
question-answer contexts. JournalofAcoustical Society ofAmerica, 77(6), 2142-2156.
99
Cooper, W. E., & Paccia-Cooper, J. (1980). Syntax and speech. Cambridge, MA:
Harvard University Press.
Cutler, A. (1977). The Context-Independence of "Intonational Meaning". Chicago
Linguistic Society (CLS 13), 104-115.
Dahan, D., Tanenhaus, M. K., & Chambers, C. G. (2002). Accent and reference
resolution in spoken-language comprehension. JournalofMemory & Language
Language, 47, 292-314.
Dilley, L. (2005) The phonetics and phonology of tonal systems, PhD thesis, MIT.
Dilley, L. and Brown, M. (2005) The RaP Labeling System, v. 1.0. Available at
http://faculty.psy.ohio-state.edu/pitt/dilley/rap-system.htm.
Dilley, L., Breen, M., Gibson, E., Bolivar, M., and Kraemer, J. (2006) A comparison of
inter-coder reliability for two systems of prosodic transcriptions: RaP (Rhythm and Pitch)
and ToBI (Tones and Break Indices). Proceedingsof the InternationalConference on
Spoken Language Processing,Pittsburgh, PA.
Eady, S. J., & Cooper, W. E. (1986). Speech intonation and focus location in matched
statements and questions. Journal of the Acoustical Society of America, 80,402-415.
Ferreira, F. (1988). Planning and timing in sentence production: The syntax-to-phonology
conversion. Unpublished dissertation, University of Massachusetts, Amherst, MA.
Ferreira, F. (1991). Effects of length and syntactic complexity on initiation times for
prepared utterances. JournalofMemory and Language, 30, 210-233.
Fox Tree, J. (1995). The effect of false starts and repetitions on the processing of
subsequent words in spontaneous speech. JournalofMemory and Language, 34, 709738.
Gee, J.P., & Grosjean, F. (1983). Performance structures. A psycholinguistic and
linguistic appraisal. Cognitive Psychology, 15, 411-458.
Gibson, E. (1998). Linguistic complexity: Locality of syntactic dependencies. Cognition,
68, 1-76.
Gordon, P. C., Hendrick, R, & Johnson, M (2001). Memory interference during language
processing. Journalof ExperimentalPsychology: Learning,Memory, and Cognition, 27,
1411-1423.
Greenberg, S., Carvey, H., & Hitchcock, L. (2002). The relation between stress accent
and pronunciation variation in spontaneous American English discourse. In Proceedings
of the ISCA Workshop on Prosody andSpeech Processing.
Grodner, D. and Gibson, E. (2005). Consequences of the serial nature of linguistic input
for sentential complexity. Cognitive Science, 29, 261-290.
Grodner, D. J., Gibson, E., & Watson, D. (2005). The influence of contextual contrast on
syntactic processing: Evidence for strong-interaction on sentence comprehension.
Cognition, 95, 275-296.
Gussenhoven, C. (1983).Testing the reality of focus domains. Language and Speech, 26,
61-80.
Gussenhoven, C. (1999). On the limits of focus projection in English. In P. Bosch & R.
van der Sandt (Eds.), Focus: Linguistic, cognitive, and computationalperspectives (pp.
100
43 -55). Cambridge, U.K.: Cambridge University Press.
Halliday, M.A.K. Intonation andgrammarin British English. The Hague: Mouton, 1967.
Hirschberg, J. (1995) Prosodic and acoustic cues to spontaneous and read speech. In
Proceedingsof InternationalCongress on Phonetic Sciences, 2, 36-43.
Ito, K. Speer, S. R. and Beckman, M. E. (2004) Informational status and pitch accent
distribution in spontaneous dialogues in English, In Proceedingsof the International
Conference on Spoken Language Processing,Nara: Japan, 279-282.
Landis, R. & Koch, G. (1977) The Measurement of Observer Agreement for Categorical
Data.Biometrics, Vol. 33, No. 1., pp. 159-174.
Kjelgaard, M., & Speer, S. (1999). Prosodic facilitation and interference in the resolution
of temporary syntactic ambiguity. JournalofMemory and Language, 40, 153-194.
Kochanski, G., Grabe, E., Coleman, J., & Rosner, B. (2005) Loudness predicts
prominence: fundamental frequency lends little. The Journalof the Acoustical Society of
America, 118 (2), 1038-1054.
Krahmer, E., & Swerts, M. (1998). Reconciling two competing views on contrastiveness.
Speech Communication,34, 391-405.
Kraljic, T. and S. E. Brennan (2005). Prosodic disambiguation of syntactic structure: For
the speaker or for the addressee? Cognitive Psychology 50: 194-231.
Ladd, D. R., Schepman, A. (2003). Sagging transitions between high accent peaks in
English: experimental evidence. JournalofPhonetics, 31, 81-112.
Lieberman, P. (1960). Some acoustic correlates of word stress in American English. The
Journalof the Acoustical Society ofAmerica, 32(4), 451-454.
Linguistic Data Consortium (1997) CALLHOME American English Speech.
http://www.ldc.upenn.edu.
Most, R. B. & Saltz, E. (1979) Information structure in sentences: New information.
Language and Speech, 22, 89-95.
Nespor, M., & Vogel, I. (1986). Prosodic phonology. Dordrecht: Foris Publications.
Ostendorf, M.F., Price P. J., Shattuck-Hufnagel S. (1995) The Boston University Radio
News Corpus. Technical Report No. ECS-95-001, Boston University.
Pierrehumbert, J., & Beckman, M. (1988). Japanese Tone Structure. Cambridge, MA:
MIT Press.
Pierrehumbert, J. & Hirschberg, J. (1990). The Meaning of Intonational Contours in the
Interpretation of Discourse. In: P. R. Cohen & J.Morgan & M. E. Pollack (eds.).
Intentions in Communication.Cambridge/MA: MIT Press, 271-311.
Pierrehumbert, J.B. (1980). The phonology andphonetics of English intonation.
Unpublished dissertation, MIT.
Pitrelli, J., Beckman, M. & Hirschberg, J. (1994) Evaluation of prosodic transcription
labeling reliability in the ToBI framework. In Proceedingsof the International
Conference on Spoken Language Processing, 123-126.
Price, P. J., Ostendorf, M., Shattuck-Hufnagel, S., & Fong, C. (1991). The use of prosody
in syntactic disambiguation. JournalofAcoustical Society ofAmerica, 90, 2956-2970.
101
Price, Patti, Mari Ostendorf, Stefanie Shattuck-Hufnagel and C.Fong. 1991. The use of
prosody in syntactic disambiguation. Journalof the Acoustical Society ofAmerica, 90:
2956-2970.
Schafer, A., Speer, S., Warren, P., & White, S. D. (2000). Intonational disambiguation in
sentence production and comprehension. JournalofPsycholinguisticResearch, 29(2):
169-182.
Selkirk, E. 0. (1984). Phonology and syntax: The relation between sound and structure.
Cambridge, MA: MIT Press.
Selkirk, E. (1995). Sentence Prosody: Intonation, Stress, and Phrasing. In: J.Goldsmith
(ed.). The Handbook ofPhonologicalTheory. Oxford: Blackwell, 550-569.
Silverman, K., Beckman, M., Pitrelli, J. Ostendorf, M. Wightman, C. Price, P.,
Pierrehumbert, J. & Hirschberg, J. (1992) ToBI: A standard for labeling English
prosody. In Proceedings of the International Conference on Spoken Language
Processing, 867-870.
Snedeker, J., & Trueswell, J. (2003). Using prosody to avoid ambiguity: Effects of
speaker awareness and referential contest. JournalofMemory and Language, 48, 103130.
Speer, S. R., Kjelgaard, M. M., & Dobroth, K. M. (1996). The influence of prosodic
structure on the resolution of temporary syntactic closure ambiguities. Journalof
PsycholinguisticResearch, 25, 247-268.
Syrdal, A. and McGory, J. (2000) Inter-transcriber reliability of ToBI prosodic labeling.
In Proceedingsof the InternationalConference on Spoken Language Processing,
Beijing: China, 235-238.
Taylor, P. (2000). Analysis and synthesis of intonation using the tilt model. Journalofthe
Acoustical Society ofAmerica. 107(3):1697-1714.
Terken and Noteboom (1987) Opposite Effects of Accentuation and Deaccentuation on
Verification Latencies for Given and New Information. Language and Cognitive
Processes2(3/4), 145-163.
Watson, D. & Gibson, E. (2004). The relationship between intonational phrasing and
syntactic structure in language production. Language and Cognitive Processes, 19, 713755.
Watson, D., Breen, M., & Gibson, E. (2006). The role of syntactic obligatoriness in the
production of intonational boundaries. Journalof ExperimentalPsychology: Learning,
Memory and Cognition, 32(5), 1045-1056.
Watson, D., Tanenhaus, M. & Gunlogson, C. (2004). Processing pitch accents:
Interpreting H* and L+H* Presented at the 17th Annual CUNY Conference on Human
Sentence Processing,Cambridge, MA.
Wheeldon, L. & Lahiri, A. (1997). Prosodic units in speech production. Journalof
Memory and Language, 37, 356-381.
Wightman, C. (2002) "ToBI or not ToBI?" In Proceedings of the International
Conference on Speech Prosody, Aix-en-Provence: France, 2002.
102
Wightman, C. W., Shattuck-Hufnagel, S., Ostendorf, M., & Price, P. (1992) Segmental
durations in the vicinity of prosodic phrase boundaries. JournalofAcoustical Society of
America, 91(3), 1707-1717.
Yoon, T., Chavarria, S., Cole, J., & Hasegawa-Johnson, M. (2004) Intertranscriber
reliability of prosodic labeling on telephone conversation using ToBI. In Proceedingsof
the InternationalConference on Spoken Language Processing.,Nara: Japan, 2729-2732.
103
Appendix A
1. The mobster paid the bounty (of thirty diamonds) to the gangster (with burly
henchmen) quickly / before the crime was committed.
2. The caterer brought the pastries (with lemon filling) to the party (for Oscar
winners) early / before the guests had arrived.
3. The gigolo sent a bouquet (of sixty roses) to the showgirl (from Hello Dolly)
on Sunday / before the performance last night.
4. The colonel assigned the mission (of killing Castro) to the soldier (with sniper
training) last night / last night at the Pentagon.
5. The wizard granted the powers (of magic healing) to the suitor (of England's
princess) last night / after being threatened with death.
6. The matriarch left the necklace (with sapphire inlay) to the daughter (of
peasant parents) secretly / before the family found out.
7. The director offered the payment (of thirty million) to the actor (of poignant
dramas) yesterday / after filming had already begun.
8. The academy presented the award (of greatest import) to the actor (of little
renown) on Sunday / last week in Los Angeles.
9. The executive delivered the statement (of corrupt actions) to the judges (of
business conduct) regretfully / before a ruling was issued.
10. The professor assigned the chapter (on local history) to the students (of social
science) yesterday / after the first midterm exam.
11. The writer pitched the story (of happy orphans) to the chairman (of Disney
Studios) at lunch / over several drinks after lunch.
12. The student gave the basket (of chocolate brownies) to the teacher (of ancient
history) today / before the start of vacation.
13. The lieutenant evacuated the soldiers (of several platoons) to a region (with
unarmed locals) yesterday / after the mysterious phone call.
14. The girl attached the posters (of missing children) to the windows (of local
buildings) today / after her shopping trip downtown.
15. The priest delivered the turkeys (with homemade stuffing) to the homeless (at
local shelters) on Thursday / before people arrived for dinner.
16. The socialite donated the suitcase (of lovely dresses) to the woman (in dirty
clothing) yesterday / after meeting her at church.
17. The lawyer left the duties (of mindless errands) to the partner (with lower
status) this morning / after the lengthy conference call.
18. The girl lent the booklet (of practice exams) to the classmate (from second
period) on Friday / before the test on Friday.
19. The gentleman sent the bouquet (of gorgeous roses) to the woman (with shiny
lipstick) on Monday / after spotting her from afar.
104
20. The millionaire assigned a chauffeur (with little patience) to his mistress (in
Southern Europe) today / after a quarrel on Friday.
21. The station offered the ballad (with minor changes) to the public (in nearby
cities) last week / after the debate last week.
22. The grandmother gave the necklace (of twenty pearls) to the grandson (from
Kansas City) on Sunday / at the annual family reunion.
23. The architect placed the statue (of Roger Sherman) in the courtyard (with
pretty flowers) carefully / with tremendous pride and satisfaction.
24. The son put his backpack (with heavy textbooks) in the kitchen (with seven
people) last night / without stopping to eat dinner.
25. The critic handed the letter (for Steven Spielberg) to the postman (with curly
sideburns) personally / in the sunshine of morning.
26. The committee allocated the money (from Tuesday's auction) to the members
(from Costa Rica) yesterday / after numerous hours of discussion.
27. The bride put the favors (of mini bouquets) on the tables (of several guests)
happily / before the wedding reception began.
28. The spy told the secrets (of deadly weapons) to the leaders (of foreign nations)
quietly / through a network of operatives.
29. The salesman conveyed his advice (on buying vases) to the clients (from rural
Texas) on Friday / after a meeting on Friday.
30. The professor assigned a project (on Asian Studies) to his students (with
heavy workloads) yesterday / without regard for other classes.
31. The tycoon lent the limo (with leather seating) to his buddies (from
Swarthmore College) often / for several days last month.
32. The referee explained the format (of soccer contests) to the players (from
Amherst College) on Friday / before the-big tournament began.
105
Appendix B
1a.
lb.
1c.
ld.
1e.
if.
1g.
Context: What happened yesterday?
Context: Who fried an omelet yesterday?
Context: What did Damon do to an omelet yesterday?
Context: What did Damon fry yesterday?
Context: Did Harry fry an omelet yesterday?
Context: Did Damon bake an omelet yesterday?
Context: Did Damon fry a chicken yesterday?
Target: No, Damon fried an omelet yesterday.
2a.
2b.
2c.
2d.
2e.
2f.
2g.
Context: What happened yesterday?
Context: Who sold her diamond yesterday?
Context: What did Megan do with her diamond yesterday?
Context: What did Megan sell yesterday?
Context: Did Jodi sell her diamond yesterday?
Context: Did Megan lose her diamond yesterday?
Context: Did Megan sell her sapphire yesterday?
Target: No, Megan sold her diamond yesterday.
3a.
3b.
3c.
3d.
3e.
3f.
3g.
Context: What happened last night?
Context: Who dried a platter last night?
Context: What did Mother do to a platter last night?
Context: What did Mother dry last night?
Context: Did Daddy dry a platter last night?
Context: Did Mother wash a platter last night?
Context: Did Mother dry a bowl last night?
Target: No, Mother dried a platter last night.
4a.
4b.
4c.
4d.
4e.
4f.
4g.
Context: What happened last night?
Context: Who read an email last night?
Context: What did Norman do with an email last night?
Context: What did Norman read last night?
Context: Did Kelly read an email last night?
Context: Did Norman write an email last night?
Context: Did Norman read a letter last night?
Target: No, Norman read an email last night.
5a.
5b.
5c.
5d.
5e.
5f.
5g.
Context: What happened this morning?
Context: Who poured a smoothie this morning?
Context: What did Lauren do with a smoothie this morning?
Context: What did Lauren pour this morning?
Context: Did Judy pour a smoothie this morning?
Context: Did Lauren drink a smoothie this morning?
Context: Did Lauren pour a cocktail this morning?
Target: No, Lauren poured a smoothie this morning.
6a.
6b.
Context: What happened this morning?
Context: Who sewed her dolly this morning?
106
6c.
6d.
6e.
6f.
6g.
Context: What didNora dohedlly this morning?
Context: What did Nora sew this morning?
Context: Did Jenny sew her dolly this morning?
Context: Did Nora rip her dolly this morning?
Context: Did Nora sew her blanket this morning?
Target: No, Nora sewed her dolly this morning.
7a.
7b.
7c.
7d.
7e.
7f.
7g.
Context: What happened on Tuesday?
Context: Who trimmed her eyebrows on Tuesday?
Context: What did Molly do to her eyebrows on Tuesday?
Context: What did Molly trim on Tuesday?
Context: Did Sarah trim her eyebrows on Tuesday?
Context: Did Molly wax her eyebrows on Tuesday?
Context: Did Molly trim her hair on Tuesday?
Target: No, Molly trimmed her eyebrows on Tuesday.
8a.
8b.
8c.
8d.
8e.
8f.
8g.
Context: What happened on Tuesday?
Context: Who burned a candle on Tuesday?
Context: What did Nolan do to a candle on Tuesday?
Context: What did Nolan burn on Tuesday?
Context: Did Steven burn a candle on Tuesday?
Context: Did Norman break a candle on Tuesday?
Context: Did Nolan burn a log on Tuesday?
Target: No, Nolan burned a candle on Tuesday.
9a.
9b.
9c.
9d.
9e.
9f.
9g.
Context: What happened last week?
Context: Who killed a termite last week?
Context: What did Logan do to a termite last week?
Context: What did Logan kill last week?
Context: Did Billy kill a termite last week?
Context: Did Logan trap a termite last week?
Context: Did Logan kill a cockroach last week?
Target: No, Logan killed a termite last week.
10a.
10b.
1oc.
10d.
10e.
10f.
10g.
Context: What happened last week?
Context: Who caught a bunny last week?
Context: What did Radar do to a bunny last week?
Context: What did Radar catch last week?
Context: Did Fido catch a bunny last week?
Context: Did Radar lick a bunny last week?
Context: Did Radar catch a squirrel last week?
Target: No, Radar caught a bunny last week.
11a.
1lb.
11c.
1Id.
1le.
Ilf.
11g.
Context:
Context:
Context:
Context:
Context:
Context:
Context:
What happened on Sunday?
Who pulled a stroller on Sunday?
What did Darren do to a stroller on Sunday?
What did Darren pull on Sunday?
Did Maggie pull a stroller on Sunday?
Did Darren push a stroller on Sunday?
Did Darren pull a sled on Sunday?
107
Target: No, Darren pulled a stroller on Sunday.
12a.
12b.
12c.
12d.
12e.
12f.
12g.
Context: What happened on Sunday?
Context: Who peeled a carrot on Sunday?
Context: What did Brandon do to a carrot on Sunday?
Context: What did Brandon peel on Sunday?
Context: Did Tommy peel a carrot on Sunday?
Context: Did Brandon eat a carrot on Sunday?
Context: Did Brandon peel a potato on Sunday?
Target: No, Brandon peeled a carrot on Sunday.
13a.
13b.
13c.
13d.
13e.
13f.
13g.
Context: What happened on Friday?
Context: Who cleaned a pillow on Friday?
Context: What did Maren do to a pillow on Friday?
Context: What did Maren clean on Friday?
Context: Did Debbie clean a pillow on Friday?
Context: Did Maren buy a pillow on Friday?
Context: Did Maren clean a rug on Friday?
Target: No, Maren cleaned a pillow on Friday.
14a.
14b.
14c.
14d.
14e.
14f.
14g.
Context: What happened on Friday?
Context: Who fooled a bully on Friday?
Context: What did Lindon do to a bully on Friday?
Context: Who did Lindon fool on Friday?
Context: Did Kelly fool a bully on Friday?
Context: Did Lindon fight a bully on Friday?
Context: Did Lindon fool a teacher on Friday?
Target: No, Lindon fooled a bully on Friday.
108
Appendix C
la.
lb.
ic.
Id.
le.
if.
1g.
Question: What happened last night?
Question: Who fed a bunny last night?
Question: What did Damon do to a bunny last night?
Question: What did Damon feed last night?
Question: Did Jenny feed a bunny last night?
Question: Did Damon pet a bunny last night?
Question: Did Damon feed a baby last night?
Response: Damon fed a bunny last night.
2a.
2b.
2c.
2d.
2e.
2g.
Question: What happened last night?
Question: Who caught a bunny last night?
Question: What did Damon do to a bunny last night?
Question: What did Damon catch last night?
Question: Did Lauren catch a bunny last night?
Question: Did Damon catch a squirrel last night?
Response: Damon caught a bunny last night.
3a.
3b.
3c.
3d.
3e.
3f.
3g.
Question: What happened last night?
Question: Who burned a candle last night?
Question: What did Damon do to a candle last night?
Question: What did Damon burn last night?
Question: Did Molly burn a candle last night?
Question: Did Damon break a candle last night?
Question: Did Damon burn a log last night?
Response: Damon burned a candle last night.
4a.
4b.
4c.
4d.
4e.
4f.
4g.
Question: What happened last night?
Question: Who cleaned a carrot last night?
Question: What did Darren do to a carrot last night?
Question: What did Darren clean last night?
Question: Did Lauren clean a carrot last night?
Question: Did Darren eat a carrot last night?
Question: Did Darren clean a chicken last night?
Response: Darren cleaned a carrot last night.
5a.
5b.
5c.
5d.
5e.
5f.
5g.
Question: What happened last night?
Question: Who peeled a carrot last night?
Question: What did Darren do to a carrot last night?
Question: What did Darren peel last night?
Question: Did Molly peel a carrot last night?
Question: Did Darren eat a carrot last night?
Question: Did Darren peel a potato last night?
Response: Darren peeled a carrot last night.
6a. Question: What happened last night?
6b. Question: Who found a diamond last night?
6c. Question: What did Darren do to a diamond last night?
109
6d. Question: What did Darren find last night?
6e. Question: Did Nora find a diamond last night?
6f. Question: Did Darren buy a diamond last night?
6g. Question: Did Darren find a ring last night?
Response: Darren found a diamond last night.
7a.
7b.
7c.
7d.
7e.
7f.
7g.
Question: What happened last night?
Question: Who sold a diamond last night?
Question: What did Darren do to a diamond last night?
Question: What did Darren sell last night?
Question: Did Jenny sell a diamond last night?
Question: Did Darren lose a diamond last night?
Question: Did Darren sell a sapphire last night?
Response: Darren sold a diamond last night.
8a.
8b.
8c.
8d.
8e.
8f.
8g.
Question: What happened last night?
Question: Who found a dollar last night?
Question: What did Jenny do to a dollar last night?
Question: What did Jenny find last night?
Question: Did Damon find a dollar last night?
Question: Did Jenny lose a dollar last night?
Question: Did Jenny find a quarter last night?
Response: Jenny found a dollar last night.
9a.
9b.
9c.
9d.
9e.
9f.
9g.
Question: What happened last night?
Question: Who sewed a dolly last night?
Question: What did Jenny do to a dolly last night?
Question: What did Jenny sew last night?
Question: Did Darren sew a dolly last night?
Question: Did Jenny rip a dolly last night?
Question: Did Jenny sew a blanket last night?
Response: Jenny sewed a dolly last night.
1Oa.
10b.
10c.
10d.
1Oe.
1Of.
10g.
Question: What happened last night?
Question: Who read an email last night?
Question: What did Jenny do to an email last night?
Question: What did Jenny read last night?
Question: Did Logan read an email last night?
Question: Did Jenny open an email last night?
Question: Did Jenny read a letter last night?
Response: Jenny read an email last night.
11 a.
1lb.
I c.
I1d.
1le.
1If.
11g.
Question: What happened last night?
Question: Who smelled a flower last night?
Question: What did Jenny do to a flower last night?
Question: What did Jenny smell last night?
Question: Did Nolan smell a flower last night?
Question: Did Jenny plant a flower last night?
Question: Did Jenny smell a skunk last night?
Response: Jenny smelled a flower last night.
110
12a.
12b.
12c.
12d.
12e.
12f.
12g.
Question: What happened last night?
Question: Who burned a letter last night?
Question: What did Lauren do to a letter last night?
Question: What did Lauren burn last night?
Question: Did Darren burn a letter last night?
Question: Did Lauren write a letter last night?
Question: Did Lauren burn a magazine last night?
Response: Lauren burned a letter last night.
13a.
13b.
13c.
13d.
13e.
13f.
13g.
Question: What happened last night?
Question: Who mailed a letter last night?
Question: What did Lauren do to a letter last night?
Question: What did Lauren mail last night?
Question: Did Logan mail a letter last night?
Question: Did Lauren open a letter last night?
Question: Did Lauren mail a package last night?
Response: Lauren mailed a letter last night.
14a.
14b.
14c.
14d.
14e.
14f.
14g.
Question: What happened last night?
Question: Who read a novel last night?
Question: What did Lauren do to a novel last night?
Question: What did Lauren read last night?
Question: Did Nolan read a novel last night?
Question: Did Lauren write a novel last night?
Question: Did Lauren read a newspaper last night?
Response: Lauren read a novel last night.
15a.
15b.
15c.
15d.
15e.
15f.
15g.
Question: What happened last night?
Question: Who fried an omelet last night?
Question: What did Lauren do to an omelet last night?
Question: What did Lauren fry last night?
Question: Did Damon fry an omelet last night?
Question: Did Lauren bake an omelet last night?
Question: Did Lauren fry a chicken last night?
Response: Lauren fried an omelet last night.
16a.
16b.
16c.
16d.
16e.
16f.
16g.
Question: What happened last night?
Question: Who peeled an onion last night?
Question: What did Logan do to an onion last night?
Question: What did Logan peel last night?
Question: Did Molly peel an onion last night?
Question: Did Logan chop an onion last night?
Question: Did Logan peel an apple last night?
Response: Logan peeled an onion last night.
17a.
17b.
17c.
17d.
Question:
Question:
Question:
Question:
What happened last night?
Who fried an onion last night?
What did Logan do to an onion last night?
What did Logan fry last night?
111
17e. Question: Did Nora fry an onion last night?
17f. Question: Did Logan chop an onion last night?
17g. Question: Did Logan fry a potato last night?
Response: Logan fried an onion last night.
18a.
18b.
18c.
18d.
18e.
18f.
18g.
Question: What happened last night?
Question: Who cleaned a pillow last night?
Question: What did Logan do to a pillow last night?
Question: What did Logan clean last night?
Question: Did Jenny clean a pillow last night?
Question: Did Logan buy a pillow last night?
Question: Did Logan clean a rug last night?
Response: Logan cleaned a pillow last night.
19a.
19b.
19c.
19d.
19e.
19f.
19g.
Question: What happened last night?
Question: Who dried a platter last night?
Question: What did Molly do to a platter last night?
Question: What did Molly dry last night?
Question: Did Logan dry a platter last night?
Question: Did Molly wash a platter last night?
Question: Did Molly dry a bowl last night?
Response: Molly dried a platter last night.
20a.
20b.
20c.
20d.
20e.
20f.
20g.
Question: What happened last night?
Question: Who sold a platter last night?
Question: What did Molly do to a platter last night?
Question: What did Molly sell last night?
Question: Did Nolan sell a platter last night?
Question: Did Molly find a platter last night?
Question: Did Molly sell a vase last night?
Response: Molly sold a platter last night.
21a.
21b.
21c.
21d.
21e.
21f.
21g.
Question: What happened last night?
Question: Who poured a smoothie last night?
Question: What did Molly do to a smoothie last night?
Question: What did Molly pour last night?
Question: Did Damon pour a smoothie last night?
Question: Did Molly drink a smoothie last night?
Question: Did Molly pour a cocktail last night?
Response: Molly poured a smoothie last night.
Question: What happened last night?
Question: Who pulled a stroller last night?
Question: What did Nolan do to a stroller last night?
Question: What did Nolan pull last night?
Question: Did Nora pull a stroller last night?
Question: Did Nolan push a stroller last night?
Question: Did Nolan pull a sled last night?
Response: Nolan pulled a stroller last night.
23a. Question: What happened last night?
22a.
22b.
22c.
22d.
22e.
22f.
22g.
112
23b.
23c.
23d.
23e.
23f.
23g.
Question: Who bought a stroller last night?
Question: What did Nolan do to a stroller last night?
Question: What did Nolan buy last night?
Question: Did Jenny buy a stroller last night?
Question: Did Nolan sell a stroller last night?
Question: Did Nolan buy a wheelbarrow last night?
Response: Nolan bought a stroller last night.
24a.
24b.
24c.
24d.
24e.
24f.
24g.
Question: What happened last night?
Question: Who sewed a sweater last night?
Question: What did Nolan do to a sweater last night?
Question: What did Nolan sew last night?
Question: Did Lauren sew a sweater last night?
Question: Did Nolan knit a sweater last night?
Question: Did Nolan sew a quilt last night?
Response: Nolan sewed a sweater last night.
25a.
25b.
25c.
25d.
25e.
25f.
25g.
Question: What happened last night?
Question: Who killed a termite last night?
Question: What did Nora do to a termite last night?
Question: What did Nora kill last night?
Question: Did Nolan kill a termite last night?
Question: Did Nora trap a termite last night?
Question: Did Nora kill a cockroach last night?
Response: Nora killed a termite last night.
26a.
26b.
26c.
26d.
26e.
26f.
26g.
Question: What happened last night?
Question: Who changed a toddler last night?
Question: What did Nora do to a toddler last night?
Question: What did Nora change last night?
Question: Did Damon change a toddler last night?
Question: Did Nora wash a toddler last night?
Question: Did Nora change a baby last night?
Response: Nora changed a toddler last night.
27a.
27b.
27c.
27d.
27e.
27f.
27g.
Question: What happened last night?
Question: Who fed a toddler last night?
Question: What did Nora do to a toddler last night?
Question: What did Nora feed last night?
Question: Did Darren feed a toddler last night?
Question: Did Nora dress a toddler last night?
Question: Did Nora feed a bunny last night?
Response: Nora fed a toddler last night.
28a.
28b.
28c.
28d.
28e.
28f.
Question: What happened last night?
Question: Who pulled a wagon last night?
Question: What did Nora do to a wagon last night?
Question: What did Nora pull last night?
Question: Did Logan pull a wagon last night?
Question: Did Nora push a wagon last night?
113
28g. Question: Did Nora pull a wheelbarrow last night?
Response: Nora pulled a wagon last night.
Download